VLA Models

TL;DR

Vision-Language-Action (VLA) models are end-to-end policies that map perception plus a language goal to robot or vehicle actions, typically built on top of a VLM backbone with an attached action head (FAST-tokenized discrete actions, a flow-matching action expert, or a diffusion-transformer chunk predictor). The filed papers disagree on which capability is the actual lever for generalist manipulation — action-pretraining-at-scale (π*0.6: a VLA That Learns From Experience (RECAP)), clean teleop data (Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models), or pointing/grounding in a unified VLM (Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models) — and at least one filed pointer (Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level) argues the VLA framing itself under-weights sensorimotor / tactile / contact dynamics.

Key claims

• Open-recipe RL fine-tuning works on flow-matching VLAs via advantage conditioning, not policy gradients. RECAP (π*0.6) sidesteps the per-token-log-likelihood problem of flow-matching action experts by binarizing a value function’s output into a “positive/negative” text token in the prefix and training the VLA with plain supervised likelihood (π*0.6: a VLA That Learns From Experience (RECAP) §IV-B, Eq. 3). RECAP reports doubling task throughput and halving failure rate vs. imitation-only baselines on the hardest tasks (π*0.6: a VLA That Learns From Experience (RECAP) §I).
• Action pretraining at scale is not strictly required if the VLM backbone has strong embodied capabilities and pointing/grounding. Embodied-R1.5-VLA attaches an action head to an 8B Qwen3-VL-derived embodied VLM and reports 97.3 LIBERO overall — matching π0.5 (96.9) and OpenVLA-OFT (97.1) which do use action pretraining, and beating the no-action-pretraining baselines by a wide margin (Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models LIBERO table). On SimplerEnv Google Robot the same model reports 92.4 overall vs π0 71.4 (Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models Google Robot table).
• Clean teleop data can be a decisive lever even at modest scale. Spirit-v1.5 reports beating π0.5 on the third-party RoboArena evaluation by curating high-quality teleop demonstrations — the data axis rather than the architecture or scale axis (Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models).
• The dominant backbone choice across open VLAs is now a Qwen-VL family model with an action head, with three filed instances: Spirit-v1.5 (Qwen3-VL backbone + DiT action head, Spirit-v1.5: A Robotic Foundation Model by Spirit AI), LingBot-VLA (Qwen2.5-VL-3B + action head, LingBot-VLA: A Pragmatic VLA Foundation Model), and Embodied-R1.5-VLA (Qwen3-VL-8B-derived embodied VLM + action head, Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models).
• NVIDIA’s open VLA stack pairs an action model (Isaac GR00T N1.6) with a video-world-model reasoning module (Cosmos Reason 2) as a deliberate bet that the VLA needs a separate world-model peer rather than a single unified model (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry). This is a structural counter-recipe to the unified-VLM approach above.
• Counter-position: the VLA framing skips the sensorimotor level — tactile sensing, contact dynamics, action policies grounded in physical interaction — and these are unlikely to come from VLM scaling alone (Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level).

Recent contributions

• [2026-11-06] Beyond Validation Loss: Coverage and Reliability Dynamics in Robot Policy Pretraining: Sun, Simchowitz & Song (Stanford REAL / CMU) add a pretraining-checkpoint-selection cell to the recipe-lever board that no existing entry targets — the board reasons about architecture, data source, RL recipe, and runtime, but treats the pretrained BC checkpoint as a given. This paper argues the pretrained checkpoint is itself a load-bearing design decision with a measurable Pareto frontier via closed-loop pass@k: pass@k peaks at different checkpoints during supervised BC pretraining (coverage early, reliability late), and checkpoints with matched pass@1 but different large-k pass@k yield different post-RL / post-SFT performance. Proposes a quantile-based offline selection criterion (predicts the closed-loop-best checkpoint without closed-loop rollouts) and JointAug, a joint observation-action data augmentation procedure that preserves large-k pass@k while maintaining or improving pass@1. Validated on MimicGen, RoboCasa VLA pretraining (action-only JointAug reaches matched SFT), and real Push-T 6-DoF wrist-only observations. Direct complement to Reinforcement Learning Enables Scaling Generalist Robot Policy Improvement‘s RL-scaling-with-task-count result by moving one step earlier in the pipeline — the checkpoint choice within a single-task BC run already determines whether downstream RL amplifies or regresses performance, orthogonal to task-diversity scaling.
• [2026-11-06] Reinforcement Learning Enables Scaling Generalist Robot Policy Improvement: Korniak, Bagatella, Römer et al. (ETH + MPI-IS + TUM + Berkeley + Microsoft + Amazon FAR + Physical Intelligence) add the first filed controlled task-count scaling study of RL post-training on a generalist robot policy — sweeping 1 → 144 training tasks on MolmoSpaces with structured scene/object held-out splits. Critic-guided recipes (OGPO, Best-of-N) show a phase transition from forgetting at low task counts to generalization at high task counts, reaching +58 pp on unseen scenes and +39 pp on unseen objects over the base policy; filtered SFT improves training tasks but transfers less reliably. Object-diversity scaling is the only axis that surpasses the base on both unseen scenes and unseen objects, and 64→144 transfer concentrates on shape-matched held-out objects (atomizers/wine bottles lifted by adding salt shakers/soap dispensers). Sharpens the RECAP (π*0.6: a VLA That Learns From Experience (RECAP)) / VINE (VINE: Taming Generative Control Policies for Reinforcement Learning) thread by localizing the generalization lift to critic-based methods and making task diversity a measured scaling axis rather than a qualitative recipe choice.
• [2026-11-06] Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs: H2RB (Park, Jung, Kim, Jang, Shin; KAIST + RLWRLD) adds a dual-interface-bridging co-training cell to the pretrained-VLA-adaptation subthread of the recipe-lever board — attack the human-robot embodiment gap at observation side (geometry-anchored alignment) and action side (decoupled-decoder tokenizer sharing an action latent space without forcing both domains through a single action mapping) simultaneously, as co-training over unpaired human + robot demonstrations. Distinct from the frozen-base residual-adaptation cells (Proxy Policy Steering: Adapting Frozen Robotic Foundation Models via Velocity-Space Proxy Residuals PPS, Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA, Learning What to Say to Your VLA: Mostly Harmless VLA Steering LFP) by actively co-training the VLA rather than wrapping it, and from the training-side representation-anchoring cells (Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment Anchor-Align, OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence DM0.5) by targeting the human-robot action-space asymmetry explicitly via a decoupled decoder rather than preserving VLM representations with a frozen copy or LR-group split. Reports real-world object transfer 8.0 → 51.8, skill transfer 62.7 in settings where co-train baselines totally fail, and OOD placement shift 6.0 → 45.3. Built on the RLDX-1 backbone; Apache 2.0 code + project page.
• [2026-11-06] cuRoboV2-Fast: Fast Torque-Limited Collision-Free Motion Planning: cuRoboV2-Fast (NVIDIA LPR; Sundaralingam, Murali, Birchfield; arXiv:2611.06913) is a planner-side contribution that tightens the LLM-as-controller row of the recipe-lever board — the classical motion-planner tool most entries on that row depend on (AGP Agent as Policy for Robotic Manipulation exposes EE-pose targets to GPT-6 Astra via an IK + motion-planner tool harness; Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness uses the same pattern; SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation SkipVLA explicitly delegates free-space transit to a classical planner so a frozen VLA is called only for contact-rich primitives) now runs at 4.4 ms mean on RTX 5090 (14.7× over cuRoboV2) and 26 ms on Jetson AGX Orin (23.2× over cuRoboV2), with Jetson Orin + cuRoboV2-Fast beating RTX 5090 + cuRoboV1 + TOPPRA on planning latency. Real YAM deployment (paired with SceneFields + GraspGenX) runs full pick-and-drop cycles at 5.93 s mean (34.6% shorter than cuRoboV2) with planning at 1% of cycle time and arm motions 45% shorter — concrete evidence that at this latency the dominant per-cycle cost shifts off the planner onto the perception front-end. At 5 kg payload, torque-limit awareness lifts success 78% → 93% vs cuRoboV1. Differentiable inverse dynamics accepts runtime payload inertia updates without kernel recompilation (load-bearing for pick tasks where carried-object dynamics change at each grasp). Companion in design-point to APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin APXinf-robo / InstinctFlash — High-Performance Serving Runtime for Robotics Models InstinctFlash on edge deployment, but on the planning side rather than the VLA serving side — and complementary to SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation SkipVLA’s “reduce VLA call count via task-structure segmentation” by reducing the latency of the classical-planner half of that segmentation.
• [2026-11-05] Perceptron Egocentric 1.0 — self-service API with SOTA hand tracking and EgoHome benchmark: Perceptron Egocentric 1.0 supplies per-hand-per-gripper action-label timelines — a stronger labeling shape than the segment-label-only annotation stacks most human-video VLA cotraining recipes consume today; enables left/right-decomposed contact structure without re-derivation downstream.
• [2026-11-05] VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding: NVIDIA VFlash provides the sharpest evidence yet that how a frozen planning VLM emits visual plans changes closed-loop policy performance, not just latency: on a Cosmos-Predict2.5-derived robot WAM cross-attending to frozen VLM features, AR’s own 2D+3D plans reach only 61.0% average success — below the 62.6% original baseline encoder and the 63.7% no-visual-plan VFlash ablation — while VFlash 2D+3D reaches 70.2%, VFlash 3D-only 69.3%. The lever is decoder-side only (frozen VLM, lightweight parallel drafter emitting primitives whose layout is declared by a tool call); action head, training budget, and VLM weights are unchanged across variants. Complements Embodied.cpp / InstinctFlash / APXinf-robo by targeting the decoding strategy rather than kernels, and adds a “visual geometry features > sub-task language features” datapoint to the frozen-VLM + cross-attention VLA composition pattern.
• [2026-11-04] Embodied Analysis — GPT-6 Astra vs Claude Opus 5 on test-tube rack placement, scored on execution directness: Embodied Analysis posts a side-by-side Astra vs Opus 5 bimanual test-tube rack placement review, with Astra ~12s direct vs Opus 5 ~19s with a mid-run correction, both scored 3/3 — another LLM-as-controller row data point, notable for the reviewer’s self-annotated caveat that two single rollouts don’t speak to reliability.
• [2026-11-04] RoboColosseum launch — open real-world robot eval initiative with 100+ robots, 140+ scenes, 10,000 free eval hours: RoboColosseum launch tweet announces an open real-world VLA evaluation fleet spanning 100+ robots (Unitree G1 humanoids, bimanual i2rt YAMs, LeRobot SO-101, Franka arms, Earth Rovers) × 140+ in-the-wild scenes, with 20,000+ rollouts already logged across two tracks (open-world zero-shot blind, fine-tune within-distribution). Observation from the launch: policy rankings vary by task and robot, reinforcing the no-single-benchmark argument that Reality Check — Pokeandwiggle's Real-World Robot Manipulation Leaderboard on an FR3 Duo Farm and RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies have been making at smaller scales. All rollouts — including failures — open-sourced on Hugging Face / LeRobot for diagnostic use, and 10,000 free eval hours are offered to academic labs.
• [2026-11-04] AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data: AIRoA-0’s inference harness runs an async rolling buffer of 2 action chunks: new chunks are denoised from noise level 1.0 to 0.1 in a single network pass, then refined to clean prediction on the next query. Two passes of the action network per query rather than the usual ten — a distillation-adjacent lever on flow-matching VLA action heads that amortizes the full denoising process across queries rather than compressing the number of denoising steps the model is trained for. Complementary to per-call speedup levers (C³ache: Accelerating World Action Models with Cross Inference Chunk Cache C³ache cross-chunk caching, TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM TurboVLA LLM-free arch) and the async-inference training-recipe thread (Real-Time Robot Execution with Masked Action Chunking REMAC, πR²: Reactive Real-time Flow Policies πR², Legato: Learning Native Continuation for Action Chunking Flow Policies Legato).
• [2026-11-04] Robotic Data Inoculation: Data Inoculation (Li, Zhou, Xu, Levine; UC Berkeley) adds a data-collection-time shared-control cell to the data-side half of the recipe-lever board distinct from the three on-robot intervention-scaling threads (Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact, MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory, Empirical results fine-tuning π0.5 on a real manufacturing task Dream Machines) and from clean-teleop scaling (Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models Spirit v1.5): a residual RL policy trained online with the same reward used for downstream policy extraction perturbs the human’s teleop command within a bounded L_∞ ball, so recovery demonstrations are elicited during initial collection against the specific failure modes a policy trained on the growing dataset would exhibit — not after iterated deploy-and-fix rounds. Downstream BC and offline RL policies on DI data match or beat matched-time baselines across three real-world precision Franka tasks (T-Cube stacking, peg insertion, keycap pulling); residual action magnitude decays during training and the residual eventually assists the teleoperator rather than perturbing. Sibling in spirit to OopsieData — a multi-lab dataset of real robot manipulation failures OopsieData (shared author Zhiyuan Zhou) but as a mechanism for producing failure data rather than a consortium aggregating it; sibling to Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning Q-Planning on the data-vs-extraction split (Q-Planning assumes failure data exists; DI generates it on-robot). Acknowledged limitation: operator stays in the loop for every episode, costing human time.
• [2026-11-04] Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation from Sensorized Human Demonstrations: Pixel-UMI (Liu, Tao, Pathak, Salakhutdinov; CMU) demonstrates a human-UMI-pretrained + DAgger-refined whole-body policy with 87.0% average task progress across five hard whole-body tasks on a research-scale wheeled humanoid. Scene-aware retargeting (EAR) + delay-calibrated action assignment + Retargeting-Error-Aware Masking (REAM) address observation-mismatch at training time; two rounds of 60 DAgger corrections per task close the remaining gap on contact-rich long-horizon behaviors. Extends the DAgger-not-RL lineage (MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory, Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact, Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics, Learning from Fleet Experience: Collective Experience Learning across 96 Robots FERT) from tabletop industrial cells to whole-body mobile on a research platform.
• [2026-11-04] Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation: Pixel (Tao, Liu, Salakhutdinov, Pathak; CMU) adds a full-stack co-designed wheeled-humanoid platform substrate row to the recipe-lever board — unified geometric-fabrics whole-body controller shared by teleoperation and policy rollouts, common training pipeline across human and robot demonstrations, hierarchical learning stack for language conditioning + long-horizon tasks, sub-$16K total BOM. Companion Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation from Sensorized Human Demonstrations Pixel-UMI carries the quantitative policy numbers; Pixel is the hardware+controller+data-pipeline substrate.
• [2026-11-04] Pranay Gupta — direct weight manipulation of π0.5 beats fine-tuning on cloth folding (50% → 68% FT → 75% weight edit): Adds a gradient-free LLM-authored weight edit on top of a frozen base VLA cell to the recipe-lever board — the first filed robotics transfer of Takehi et al.’s Weight-Space Programming paradigm (Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights) onto π0.5. On cloth folding evaluated on RoboDojo, reports Base 50% → FT on Astra corrections 68% → direct weight manipulation 75% — a 25 pp lift over base with (presumably) no gradients, and the 7 pp gap over fine-tuning is the first filed datapoint that LLM-authored weight edits can compete with gradient-based fine-tuning on a flow-matching action-expert VLA. Direct counter-pattern to Empirical results fine-tuning π0.5 on a real manufacturing task‘s finding that LoRA underperforms full FT on π0.5 by 23–43 pp (“new embodiment is not a low-rank edit”); reconciliation hinges on whether cloth folding is specifically obstruction-removable per Takehi’s Qwen vs Gemma ability-vs-obstruction dichotomy. Tweet-only; sample size, π0.5 checkpoint, task subset, and forward-pass budget unreported — flagged for /bud refresh on the promised full breakdown.
• [2026-11-03] Democratizing Robot-Use Agents: A Unified Harness with Open Models: uni-robot-agent (Yuan, Zuo, Huang et al.; Berkeley + Tsinghua) adds a democratized-open-weight-RUA + unified-harness cell to the LLM-as-controller row of the recipe-lever board — out-of-the-box Qwen3.8-Flash-Next under the harness scores 31.39 on RoboDojo vs 28.97 for the published GPT-6 Astra-RoboProbe result and beats “many VLA/WAM baselines” on the same substrate, with competitive performance on 19 real-world tasks. Direct counter-datapoint to Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy Astranomy’s capability-cliff finding — argues the open-weight controller + right-harness combination is enough to clear the Astra bar on RoboDojo without privileged inputs. Sibling to OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA and RobotUse: Allocating Computation, Context, and Decisions RobotUse on the same LLM-as-controller cell, but swaps a frontier closed model for an open-weight one and makes the harness training-ready for joint model + harness optimization.
• [2026-11-03] BiDex-100: A Comprehensive Benchmark for Bimanual Dexterous Manipulation Across 100+ Simulation Tasks: BiDex-100 (Wang, Zheng, Zhu, Xu, Shen, Cai) adds the first filed dexterous-bimanual, capability-axis-factored, paired-OOD evaluation substrate to the recipe-lever board — 106 tasks across generalization / instruction-following / limited-demonstrations / long-horizon / memory with two high-DoF hand embodiments, 8,080 success-validated teleop demos (6.25M frames), and seven VLA baselines evaluated under one unified protocol. Fills the gap between general-manipulation sim benches (RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies RoboDojo, LIBERO, RoboTwin, RoboCasa) and the dexterous-policy lineage (DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations DONIC, The Cartesian Hand: In-Hand Manipulation with All-Linear Fingers) with a measurement surface that scores robustness directly via per-task ID↔OOD pairs rather than inferring it from aggregated benchmark deltas. Sibling methodology to The Imitator Game: Benchmarking Robot Imitative Ability Beyond Action Prediction The Imitator Game on the “level-grade the axis rather than aggregate” axis. Per-task / per-axis / per-model numbers and VLA baseline identities not retrievable from abstract at filing time.
• [2026-11-02] 3D World Simulator: Hybrid Particle-Voxel Dynamics for Multimodal 3D World Simulation: 3D World Simulator (§6) demonstrates an explicit-3D learned WFM as a policy-in-the-model evaluator: policies are run inside the model by rendering the camera views they expect, feeding actions back, and continuing autoregressive rollout to distinguish success from failure on Push-T, cloth folding, and plug insertion. Adjacent to WMBench (GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation) and the “WFM as rollout backend” cell of the VLA recipe-lever board, but with explicit 3D geometry and co-located tactile predicted at every step rather than 2D video tokens.
• [2026-10-30] Kalyan — dense Argus-format annotations for >5,000 hours of public robot data, labeled by GPT-6.1 Sol: Kalyan’s dense-annotation release stakes a hierarchical-factorization position adjacent to AGP (Agent as Policy for Robotic Manipulation) and Sunday Robotics ACT-2: VLMs encode task intent, the action head still absorbs the complex sensorimotor mapping, so in the scarce-robot-data regime the lever is to let the VLM own memory + planning and break tasks into fine-grained steps so the action head reduces to a small set of atomic skills seen across many more scenes/objects/embodiments. Cites VLM4VLA (arxiv:2601.03309) and LAP (arxiv:2602.10556) as prior negative evidence that naive VLM backbone scaling alone does not translate to VLA success; the released >5k-hour dense per-episode subgoal + outcome corpus is pitched as the public data substrate the hierarchical recipe needs.
• [2026-10-30] DexHoldem: An Agentic Robotics Benchmark for Dexterous Manipulation in Texas Hold'em: DexHoldem (Chen, Chu, Sun, Zhou, Xu, Gao, Zhai, Yang, Ma; HKU + Berkeley) adds a cards-and-chips fine-motor policy bench row with a clean pretrained-vs-from-scratch gap: π0.5 and π0 tie at 38 SP / 80 rollouts and RDT reaches 24 SP, while from-scratch DP-Transformer / RDT-small / ACT / BAKU / DP-UNet cluster at 1–11 SP — a new discrete-state verifier regime (poker game state is enumerable per frame) that complements the existing Reality Check / ArmnetBench substrate on primitives neither covers (14 card and chip primitives, 1,470 teleop demos). Also stakes out an agentic-perception axis on the same hardware: 36 snapshot problems where a frontier VLM under a coding-agent harness (Claude Code / Gemini CLI / Codex) must emit an 8-field structured game state — Codex + GPT 5.5 wins field-wise average (66.8) while Claude Code + Opus 4.7 wins strict problem-level exact-match (34.3), and GPT 5.4 mini + Claude Haiku 4.5 both score 0.0 on showdown-outcome parsing (terminal-event parsing is a distinct capability from mid-hand state recovery). First filed benchmark where the frontier-VLM-as-perceiver + VLA-as-executor factorization is scored on both halves independently against a discrete ground truth.
• [2026-10-29] ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning: ThinkWorld (Chaehyun Kim) adds a predicted-future-matches-observed-outcome RL cell to the recipe-lever board distinct from RECAP’s advantage-conditioning on binarized value tokens (π*0.6: a VLA That Learns From Experience (RECAP)), VINE’s end-to-end BPTT through flow-matching (VINE: Taming Generative Control Policies for Reinforcement Learning), Q-Planning’s frozen-policy + off-policy Q-function reweighting (Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning), and WVM’s separate world value model (World Value Models for Robotic Manipulation): the reward signal is observation-vs-predicted-future match on CoD proposals, which turns action-free human videos into first-class training data under the same objective. Positions it as the first filed VLA RL recipe where the per-step reward does not require any action-label channel, only the observed-next-observation channel both human and robot streams naturally carry.
• [2026-10-29] CausGround: Causal Interaction Grounding for Adaptive Embodied Planning: CausGround (Shen, He, Guo et al.) adds a non-parametric causal-knowledge-library + frozen-MLLM-at-inference cell to the recipe-lever board distinct from Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents RPG (symbolic skill library + system prompt), Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement COAP (code library as the policy), and OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA (zero-abstraction ROS 2 CLI): interaction experience is distilled into reusable action preconditions and effects over a factorized world-state representation of object attributes + inter-object relations, and the planner grounds task-relevant causal dependencies in the current observation to run backward reasoning over preconditions + forward reasoning over effects. Reports 24.9% → 53.3% on adapted VirtualHome, 43.0% → 57.0% closed-loop on adapted RoboCasa, 25.8% → 31.9% on SWITCH interaction reasoning, and real-world sequential manipulation driven backward from an MLLM-interpreted goal — all with no parameter updates. Sits one layer above the manipulation-primitive-focused neighbors on the recipe-lever board: targets the high-level planning + interaction-reasoning tier that MLLMs currently fail at (per the SWITCH baselines), rather than the sensorimotor execution tier where the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level counter-position lives.
• [2026-10-28] Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents: RPG (Berkeley; Abbeel/Haghtalab/Sastry labs) adds a no-weight-update, no-action-head, frontier-LLM-coordinates-perception-and-control row to the recipe-lever board — the shipped artifact is a system prompt + symbolic skill library, with a multimodal LLM at test time binding them to real hardware after a common calibration pass. On 22 held-out initializations, 95.0% success beats ASPIRE: Agentic /Skills Discovery for Robotics ASPIRE’s 75.5% and CaP-Agent0 (GPT-6 Astra Pro) 60.0%; 30/30 real-robot trials across three tasks after calibration. Direct successor-cell to ASPIRE / Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement COAP / OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA on the “no learned policy, library + orchestrator is the policy” axis — and distinct from HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own HomeBody-style typed-skill-call systems by growing the library through automated practice rather than hand-authoring the five skills.
• [2026-10-28] Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement: COAP (Hu, Hu, Hong, Chen, Liu; NTU S-Lab) stakes out the sharpest no-model-at-test-time cell on the recipe-lever board — reframes the embodied world as an Embodied Turing Machine (tape = state, rules = policy) and argues that if state can be tracked accurately from camera + proprioception, decision-making is written entirely in code with neither a VLA nor an Agent Harness in the loop. Reports 70.24% success on RoboDojo’s 42 bimanual tasks with no model at test time — directly comparable to RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies‘s 30-policy sweep and to InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data InternW0-Δ’s 23.9% on the same bench. Closest neighbor is OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA (zero-abstraction, Claude Code + ROS 2 CLI only) but moves one step further by removing the LLM at inference entirely — all intelligence shifts into a pre-authored shared library coding agents develop in a closed loop. The three advantages named — Explicit State (variables in code, not weights/context), Execution (controllable, deterministic, cheap/fast, flexible failure recovery), Extensibility (new tasks reuse/inherit/extend one shared library) — are proposed as the operational advantages of code as a substrate over VLAs and Agent Harnesses. Positions COAP as also a data engine for downstream VLAs/Agent Harnesses because the library is episode-general rather than per-instance. The paper’s own stated upper bound: accuracy of state representation + code-logic robustness, not model capacity.
• [2026-10-27] AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra: AHa-3D is a frontier-agent-as-controller datapoint on a non-robot payload: GPT-6 Astra orchestrates a 3D vision tool library (Pi3X, SAM3, GVHMR, Kimodo) to produce editable Blender scenes with human motion, in direct architectural neighborhood to the “GPT-6 Astra as native robot policy” probes (Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy, Has GPT-6-Astra Achieved Embodied Intelligence?) but with video re-rendering + action editing instead of real-robot control.
• [2026-10-26] Recova: Agent-Guided Failure Recovery for Autonomous Robotic Manipulation: Recova (NVIDIA + UCSD + UT Austin; Liu, Cheng, Bjorck, Yu, Yin, Kautz, Fan, Zhu, Sifei Liu) adds a twin-based co-development of task + recovery policy with agentic monitor-invoke-verify-resume cell to the recipe-lever board distinct from EgoRecovery (off-robot egocentric capture + gated corrective-intent), Enact / Dream Machines / Robo Robotics (on-robot DAgger with human corrections only), and FailSafe (companion VLM emitting recovery deltas against a frozen VLA at runtime): one central agent diagnoses failures and writes/tests corrective programs in a reconstructed digital twin, collects successful task and recovery rollouts for two separate policies, and at deployment monitors progress → invokes a learned-or-programmatic recovery → verifies scene restoration → resumes task execution, with human demonstrations resolving failures that no recovery handles and feeding back into the recovery dataset. Reports 78.8% / 64.9% mean success across six LIBERO-Pro settings / four MolmoSpaces categories vs 71.7% / 38.0% strongest baselines; parallel collection across four real-robot workstations lifts mean success 23.8% → 77.5% via DAgger fine-tuning, with recovery-skill invocation raising it further to 87.5%. Sharpest filed datapoint on the capabilities-compound-over-rounds axis: on one task over four collection rounds, human-intervention rate falls 87.5% → 0% — concrete evidence that twin-authored recoveries progressively eliminate the human-in-the-loop step. First filed instance putting twin-based diagnosis + corrective-program authoring + twin-authored training data for both policies + agentic deployment-time orchestration + real-world DAgger that routes rollouts to the matching policy in one loop.
• [2026-10-26] Agentic World Action Models: AW-0 (Liu, Niu, Zhang et al.; Berkeley + NVIDIA GEAR + Stanford + UT Austin + Fudan + collaborators) adds a joint-denoised code+video+action MoT with hybrid executable-plan fallback cell to the recipe-lever board — one Mixture-of-Transformers with three experts (code = LLaDA diffusion LM + frozen Qwen3-VL ViT; video = Cosmos-Predict2.5 DiT; action = separate DiT) communicating via masked joint attention in blockwise-causal order code → video → action. Programs are both training-time conditioning that shapes the predicted future and deployment-time executable code; by default the action chunk drives the robot while the generated code conditions it, and in hybrid mode the decoded slice is executed Code-as-Policies-style against perception/control primitives with the action chunk as fallback on validity-check failure. Pretrained on 2,000 h YAM bimanual teleoperation + 1,000 h egocentric human manipulation under a 50-skill library (23 YAM + 27 egocentric) distilled by Gemini 3.1 Pro + Claude Opus 4.8 from one episode per task, with a cached subtask→program mapping reused on contact/motion-pattern matches. Reports 50.7% average on RoboTwin 2.0 (80.8% clean, 29.2% random in hybrid mode) and 71.1% average on 8 real-world bimanual tasks on YAM parallel grippers + Tianji-Sharpa 22-DoF dexterous hands vs 46.0% DreamZero, 43.6% π0.5, 39.9% WLA, 35.9% EgoScale — largest absolute gains on Flip Scoreboard (+28), Pack Backpack (+15), Solve Cube (+20), and Arrange Fruits (+21). Load-bearing ablations: C → V → A ordering beats V → C → A on 7/8 tasks; a dedicated Qwen3-VL ViT for the code expert beats reusing the video expert’s Wan2.2 VAE features on all 8 tasks (71.1 vs 55.6 avg, +25 pp on Pack Backpack); joint denoising beats autoregressive teacher-forced code → video → action on 7/8 tasks (71.1 vs 54.5 avg); full annotated programs with comments beat comment-free code on all 8 tasks and natural-language subgoals on 7/8; removing skill calls hurts on all 8 tasks (Deal Poker 58.3 → 42.0). Hybrid-mode on Fold and Stack Towels reaches 90.6% vs 84.4% for pure action-chunk execution with the same checkpoint — concrete answer that dexterous folding stays on the action expert while structured approach + stacking motion benefits from executed code. Sibling to Playful Agentic Robot Learning RATs’s play-stage frozen-library pattern but freezes against demonstration data and marries the library to a jointly-trained WAM, and sibling to Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA on the “wrap frozen policy with code-authored orchestration” axis but with the orchestration integrated into training rather than inference-time only.
• [2026-10-23] Toward Recursive Self-Improvement in Physical AI: Luo et al. physical-RSI survey (NUS + HUST + SJTU + NTU + HKBU) provides a vocabulary for the “self-improving VLA” subthread of the recipe-lever board — distinguishing policy updates (behavior directly), world-model updates, memory/skill/code retention, and learning-procedure updates as four separable targets of retained change under a controlled frozen-knowledge comparison. Relevant to the self-improvement-via-DAgger / human-in-the-loop line (Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact, MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory, Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics, Empirical results fine-tuning π0.5 on a real manufacturing task, Learning from Fleet Experience: Collective Experience Learning across 96 Robots CEL/FERT), to the Q-function-only self-improvement line (Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning), and to RECAP’s advantage-conditioning recipe (π*0.6: a VLA That Learns From Experience (RECAP)) — none of which currently distinguish cumulative single-loop improvement from recursion under the paper’s two-round frozen-knowledge test.
• [2026-10-22] CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation: CW-WAM (anonymous; cw-wam-project.github.io) adds a joint wrench prediction + outer-loop admittance controller row to the recipe-lever board on the WAM side — distinct from the policy-architecture answers of FACT: Force-Aware Contact-Rich Manipulation via Timestep Modulation — demystifying when and why VLAs fail on contact-rich tasks (per-layer force modulation of a VLA), ReForce — force-aware retargeting via closed-loop residual force tracker ReForce (sim-trained residual force tracker atop an ACT reference), Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction T-Rex (variable-rate MoT + tactile VQ-VAE on kHz robot-side touch), and the data-side pretraining-corpus answer of FACT: A Multimodal Force-Action-Track Foundation Model for Human-to-Robot Dexterity FACT (force in human pretraining via EgoForce): CW-WAM keeps the WAM factorization and makes the predicted 6-D contact wrench a first-class modeling target paired with motion via pair-causal attention, with a long fixed-length wrench memory preserving sustained interaction history. The predicted wrench chunk becomes the reference for a 20 Hz admittance controller inside a slow–fast control system (WAM chunk-rate → 20 Hz admittance → 1 kHz robot controller), so force tracking does not stall between WAM queries. On four contact-rich tasks, square peg insertion (38 mm peg, 40 mm hole, 1 mm clearance per side, ~50 mm depth) lifts 25% → 70% and power adapter seating 40% → 95% vs the no-wrench WAM baseline; force-reference RMSE drops 4.8 N → 2.7 N on peg insertion and 2.0 N → 1.3 N on wiping with smaller force fluctuations during successful execution. Lifting-height accuracy also improves across variable loads (30 g vs 830 g identical-looking boxes), concrete evidence that the wrench channel resolves ambiguities vision alone cannot. Sharpest filed WAM-side realization of the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position, and lands on the same tight-clearance peg-insertion regime where Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts stakes the proprioception-only counter-position — enabling a direct wrench-aware-WAM vs blind-proprio-RL head-to-head on this class of tasks.
• [2026-10-22] Interactive World Simulation with Cosmos for Physical AI: Cosmos Simulation deliberately scopes closed-loop policy utility out of the report (“policy-driven simulator utility requiring separate closed-loop validation” [§1]) and positions the WFM as the environment a downstream policy interacts with, not as the policy itself. Complements rather than competes with the WFM-as-policy line (Dyna-2, FastWAM, Long-WAM, LingBot-VA) by productizing the simulator-as-training/evaluation-substrate deployment pattern at frontier open-model scale across robot manipulation, controllable video transfer, and multi-view AV + LiDAR.
• [2026-10-21] Learning from Fleet Experience: Collective Experience Learning across 96 Robots: CEL/FERT (Tsinghua + AgiBot + Infinigence AI) adds a fleet-scale per-task self-improvement with binary labels only cell to the recipe-lever board, distinct from the single-robot DAgger service lineage (Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact, MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory, Empirical results fine-tuning π0.5 on a real manufacturing task) and from the value-function reweighting answer (Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning Q-Planning). After SFT + (M)HG-DAgger initialization (which lifts mean success 36.3% → 81.0% across five bimanual tasks), 96 AgiBot G2 robots run autonomous rollouts; humans supply only binary success/failure labels and resets. FERT uses failure-trajectory observations as queries against a shared successful-episode buffer, upweights visually/proprioceptively-similar successful segments, and trains plain online BC on successes — no RL, no value head, no corrective-action signal. Mean success 97.3% across the five tasks (+16.3 pp over init, 86.0% failure reduction). Phone packaging reaches 98.3% / 92% failure reduction at 96 robots, 60.0 → 97.5% on phone charging (93.8% failure reduction). Fleet-size scaling at matched per-robot collection time: 4 robots 50.8% (below the 79.2% starting policy), 16 robots 86.7%, 96 robots 98.3% — concrete calibration for the minimum fleet size at which the self-improvement loop pays. Ablation isolates two load-bearing design choices: removing the ±30-frame context around retrieved segments drops success 86.7% → 57.5% at 16 robots (retrieval alone is insufficient — temporal continuity matters), and policy-sync cadence (every 500 optimizer steps) + priority retention (88% per sync) are the empirically-maximizing settings for the asynchronous training staleness–stability trade-off. Deployment-time evaluation on two robots held out from training reaches 93.3–99.2% success across the five tasks (60 trials × 2 robots per task). FATS infrastructure layer (HotSwarm fresh-identity restart + SwarmLink pooled-observation + prioritized interaction traffic) absorbs 30 fault incidents affecting 22 robots at 1.3% total-robot-time loss during a 96-robot / 9,834-episode run — operational-side evidence that fleet-RL/DAgger infrastructure can run at this scale without hand-rolled plumbing per deployment.
• [2026-10-21] R²-WAM: Data-Free Continual Learning for World Action Model with Rectified Generative Replay: R²-WAM (Fang, Liu, Shi, Zhou, Zhu, Guo, Song, Gao; UESTC + Shanghai AI Lab + Fudan + SJTU + Tongji + SII) adds a data-free continual-learning row to the recipe-lever board distinct from action-pretraining / clean-teleop / unified-VLM-pointing / frozen-WFM+action-expert / symmetric-I/O-backbone / LLM-as-controller — the WAM regenerates rollouts of previously learned tasks from language instructions alone, rectifies them with an agentic VLM evaluator (truncate at task completion, discard non-successful) and an optical-flow physical-consistency filter (robot-only sim vs 3D-tracker flow on generated frames), and replays them against sequential fine-tuning. Reports LIBERO 4-suite 92.7% SR / 2.33 NBT with 20 rectified rollouts per prior task, within 1.9 points of real-data ER (94.6%), vs sequential fine-tuning 12.5%. Raw generative replay (no rectification) is 70.4%; ablation sharpens the lever: w/o truncation 70.8%, w/o filtering 80.8%, both 94.0% on LIBERO-Spatial — the two rectifications are jointly necessary and individually insufficient. Real-world 5-task sandwich-making sequence on arm + dexterous hand: 89.5% vs ER 91.8% vs Seq-FT 28.5%. Load-bearing mechanism finding via teacher-forced probe: forgetting concentrates in the WAM’s inverse-dynamics channel while video prediction on prior tasks stays visually plausible — a forgotten WAM still imagines the past correctly, which is what makes self-regeneration viable. Classical CL baselines (EWC 13.3%, LwF 14.9%, Seq-LoRA 11.1%) do not materially improve on Seq-FT in this setting. Sibling in framing to the WAM lineage (Causal World Modeling for Robot Control (LingBot-VA) LingBot-VA, Long-WAM: Scaling the Context of World-Action Models Long-WAM, InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data InternW0-Δ) but staking the “weights-only deployment, no prior-task data available” cell none of them targets, and sibling to EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data EgoWAM on the “which channel transfers what signal” ablation axis.
• [2026-10-20] Flora: An Open Model Suite for Scaling Robotics Data Engines: Flora (NVIDIA GEAR) is the open-weight data engine behind GR00T training — covering the three VLA-pipeline label types (timestamped captions, dense process rewards, outcome rubrics) in one open-weight suite deployed at 200K hours/day. Direct answer to the subtask-labeling-at-scale bottleneck π0.5 / RT-H / SARM identified for VLA pretraining; now shipped as a production-scale open artifact rather than a prompting recipe.
• [2026-10-19] FACT: A Multimodal Force-Action-Track Foundation Model for Human-to-Robot Dexterity: FACT (Stanford; Cao, Hong, Yang, Choi, Fei-Fei Li, Wu, Huang Huang) adds a two-stage planner-plus-force-refiner with force-aware human pretraining row to the recipe-lever board distinct from VLM-with-action-head (π*0.6, Embodied-R1.5), frozen-WFM+action-expert (µ₀), unified causal AR flow-matching (LingBot-VA), and symmetric-I/O backbone (Rho-1): a Track-Guided Coarse Planner (actions + future video + 2D visual tracks) is coupled with a Force-Action Joint Refiner that predicts future force and uses it to correct the coarse action. Pretrained on EgoForce, a large-scale egocentric human manipulation corpus with recorded contact forces (released with the paper). Reports up to +60 pp success rate over T-Rex on ten contact-rich tasks (tool use, musical instruments, precision lab procedures) with only 10 robot + 90 human trajectories at fine-tuning — direct empirical support for the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position that VLA scaling under-weights force/contact, now with the sharpest filed quantitative answer at foundation scale. Load-bearing ablation: force supervision in the human pretraining data carries the downstream robot gain, not architecture tweaks or video co-training alone. Human:robot ratio sweep shows FACT learns from human demonstrations nearly as effectively as from robot demonstrations — a quantitative exchange-rate measurement for the human-data-as-pretraining-substrate recipes already on the board (EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale, Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2, The Will to Act — Natura-1 Natura-1). Sibling in force-integration to FACT: Force-Aware Contact-Rich Manipulation via Timestep Modulation — demystifying when and why VLAs fail on contact-rich tasks (same-acronym FACT, per-layer modulation of the VLA by current force; integration lives in the policy architecture) and ReForce — force-aware retargeting via closed-loop residual force tracker ReForce (sim-trained residual force tracker atop a reference) — FACT (this paper) moves the lever one layer upstream into the human pretraining corpus.
• [2026-10-19] Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts: Bronars, Park, Kedia & Agrawal (MIT) add the sharpest proprioception-only, no demos, no F/T, no vision, no tactile cell to the recipe-lever board — blind policies observing only short joint-position/velocity/action history, trained purely in simulation with PPO over randomized parts/grasps/poses, hit 73% zero-shot real-robot peg insertion across 8 pegs (incl. unseen tighter 0.5 mm clearance), with matched visual-policy controls in sim generalizing worse to unseen part shapes/sizes than blind. Sibling to A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control SGS on the pure-sim-to-real-distillation-for-NIST-assembly cell and to FLASH: Fast Learning via GPU-Accelerated Simulation for High-Fidelity Deformable Manipulation in Minutes FLASH on the “specialized sim engine + no real data” lineage, but drops vision entirely rather than keeping it as a student input. Vision costs less when it works (0.44 s vs 0.76 s median) but overfits to appearance on unseen parts; linear probes on the blind-policy latent fit in sim and used unchanged on real robots recover task-relevant features — argues the policy reasons through contact rather than finding an open-loop mechanical pattern. Open-sources mjfit, a procedural MuJoCo asset generator from exact convex pieces that preserves tight clearances (V-HACD/CoACD fill them; MuJoCo SDF is too slow) — concrete tooling for the sim-side of this and other tight-clearance assembly recipes.
• [2026-10-19] Has GPT-6-Astra Achieved Embodied Intelligence?: SenseNova Joint Research Team runs the breadth-axis companion to Astranomy — 12 benchmarks × 8 real robot configurations × multiple embodiments, with a System-2 Astra + System-1 π0.5 close-loop as the agentic-harness recipe, and formally separates embodied competence (success under unbounded deliberation + ICL demos) from practical embodied intelligence (success under a running world clock + no in-domain demos).
• [2026-10-18] FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs: FoveaVLA (Stanford TML; Chen, Zha, Karen Liu, Tevet) adds a second replace-the-wrist-camera-with-learned-foveation cell to the recipe-lever board next to EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras EyeRobot 2.0 — but keeps cameras fixed and treats crop coordinates themselves as the action, decoded from π0.5’s inherited PaliGemma location-token vocabulary (no new policy head, no RL pipeline, no mechanical eye). Alongside each flow-matched action chunk the model emits one 2D interaction-region box per external view, supervised by RoboInter annotations on DROID targeting the next policy query so the predicted region accounts for gripper/object motion during chunk execution; box tokens decode from the cached prefix in parallel with execution so there is no added wait at chunk boundaries. Two configurations on top of π0.5 — Single (one external RGB-D view) 70.8% and Dual (two external views) 77.1% vs wrist-equipped π0.5 52.1% averaged over four real-world DROID suites, matching the wrist baseline on plain pick-and-place and improving on distractor-heavy, cloth, and far-away-object tasks. Patch-token budget argument makes the lever concrete: a resized 1280×720 full view covers an interaction region with ~15 of 256 ViT patch tokens, vs ~256 of 256 for a same-sized crop at the original resolution. Second large-sample datapoint on “decouple vision from action is the architectural answer to Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level Malik’s sensorimotor counter-position” after EyeRobot 2.0’s 67% vs 52% on seven tasks — two independent recipes (gaze-as-action vs crop-coord-as-action) now converging on the same “wrist camera is not the right VLA perception primitive” conclusion with Karen Liu involvement on both.
• [2026-10-16] Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy: Astranomy (NVIDIA SRL; Yang, Khetan, Zhang) runs the first filed controlled, large-sample characterization of the LLM-as-robot-policy cell on the recipe-lever board against the same RoboLab-120 substrate used by other filed VLA/WAM entries — three fixed policy forms (Direct API, Agentic Codex session, Orchestrated review-of-Cosmos3-Nano proposals), 5 seeds × 120 tasks, with inference-time ablations across action representation (EEF > joint > B-spline), chunk horizon (short beats long by 6pp), geometry prompts (load-bearing for Direct, 2.7pp for Agentic), context history (text > visual-only, 100% on 23-task subset with full history), and model tier (Astra 72.7% / Sol 33.3% / Luna 2.0% on 30-task Direct). Agentic Policy tops RoboLab-120 at 74.8% — directly comparable to RobotUse: Allocating Computation, Context, and Decisions RobotUse’s 45% on the same benchmark and ARC: A Reasoning Recipe for Robot Foundation Models ARC’s +pp wrapper of π0.5/Cosmos3-Nano, giving the first filed common-axis read of LLM-as-policy vs reasoning-wrapper-over-VLA/WAM. Two load-bearing structural findings: (i) persistent conversation history is the dominant lever (Direct 72.7% → Agentic 86.0% on the 30-task suite), while agentic tool-use (19.7 tool-calls/episode average: stereo triangulation, IK, image crops) does not clearly improve over the no-tools Agentic baseline — argues the Agentic lift is memory-driven, not tool-driven; (ii) the two-order-of-magnitude cost premium ($36.40 / 30.9M tokens per successful Agentic episode vs sub-second specialized VLA inference) is the sharpest filed quantitative weight behind Robot-Use Agents Isola’s “robot-use agents” positional thesis that per-robot performance stays below dedicated policies while intelligence diffuses through the ecosystem. Inference accounts for 94.9–97.8% of estimated episode duration across all three Astra configurations (18.56 s/call Direct EEF) vs 4.6–25.6% for FLUX 3 Action / Cosmos3-Nano-Policy / π0.5 reference baselines — a measured per-episode-cost / per-chunk-latency Pareto that the Robocurve Comparing Fable 5.1 to Fable 5 on robotic manipulation / Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK small-N threads had been circling without large-sample weight. §6 failure analysis: dominant modes are contact establishment (gripper repeatedly adjusts around an object without securing) and placement completion, directly reproducing the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position prediction at RoboLab-120 scale rather than qualitative demo scale. Orchestrated Policy intervenes on 14.5% of simulation steps with 64.4% of intervention blocks addressing grasping and 20.2% addressing object-selection/planning — the inference-time inverse of ARC (Astra reviewing Cosmos3-Nano proposals rather than Cosmos consuming Astra reasoning traces).
• [2026-10-16] Introducing Glimpse: Glimpse demos as System-2 planner for the Physical Intelligence π0.5 VLA on “put cream cheese in bowl, then put bowl on top of cabinet”: both Glimpse-planner and GPT-6-Astra-planner complete the task, but Glimpse costs $0.035/episode vs Astra's$0.249/episode (7.1× cheaper). Elorian flags direct action outputs for robots as roadmap — current build is understanding/planning only.
• [2026-10-16] Residual Modeling Closes the Regression and Generative Policy Gap in Robot Learning: HT-Policy (Zhou, You, Wan, Dong, Gao, Mao; Tsinghua + MIT) adds a loss-side row to the recipe-lever board distinct from the architectural / interface / RL / data levers already cataloged: swap MSE regression for heteroscedastic Student-t (state-dependent scale head + Student-t noise model) on the same flow-matching backbones (GR00T N1.7, π0.5, Cosmos3-Nano) and recover competitive task success at 1.82× / 1.88× / 17.83× lower inference cost. Direct challenge to the folk “flow matches action multimodality” explanation on this page: analyzes real-world robot demonstration residuals and shows substantial state-dependent scale variation + heavy tails, with MSE’s gradient allocating disproportionate magnitude to high-residual observations — the fix is at the loss, not the output distribution. Task-success parity reported as GR00T N1.7/RoboCasa-GR1 50.0% vs 42.3% Flow, π0.5/LIBERO 97.0% vs 97.1%, Cosmos3-Nano/LIBERO-10 96.5% vs 95.6%. Sibling in framing to Behavioral Cloning Mystery Behavioral Cloning Mystery (same “diagnose BC pathologies at the objective level without perception as confounder” shape) and loss-side counterpart to B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations B-spline Policy’s output-space lever. The 17.83× speedup on Cosmos3-Nano is the sharpest datapoint on this page for backbones where flow-matching denoising is the inference bottleneck.
• [2026-10-15] RoboSimStudio: Smartphone Video to Multi-Physics Simulation via Coding Agents: RoboSimStudio feeds its coding-agent-authored multi-physics simulations into a downstream Graph-as-Policy (GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks) rather than a monolithic VLA — same Berkeley/NVIDIA lab lineage — reinforcing the GaP position that graph-structured policies are a viable complement to end-to-end VLAs, now with an upstream coding-agent-authored simulator providing the training distribution.
• [2026-10-15] HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution: HumanoidToolBench (SNU-PI, Jang et al.) adds the first filed real-humanoid tool-selection ↔ mobile-execution diagnostic to the recipe-lever board — 18 tasks × 3 scenarios × 3 execution levels × 2 tool-set modes on Unitree G1 with a five-event sub-success chain that makes the selection-vs-execution gap directly readable. FastWAM (6B WAM) leads 6-of-9 standard-mode cells and GR00T N1.7 leads selection (L0 Select 77); both collapse at the “tool-target contact → success” transition (~87% correct-first-contact but ~37–53% final success) and lose 10–20 pp in decoy mode. Two load-bearing probes reinforce prior theses: standard-only GR00T N1.7 drops BallMove correct-first-contact 85→74% / BallRetrieve 59→36% on unseen tools, and still breaks the ice 84% of the time under the instruction “The weather is nice today” (vs 77% with the correct instruction) — sharp empirical datapoint for the Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models visual-shortcut thesis. ToolBook corpus (3,003 sim + 91 real teleop demos via Meta Quest, 55 tool meshes, CC BY-NC 4.0) and all code/weights released under permissive licenses.
• [2026-10-13] Gemini 4 Argon for Robotics: Google’s first-party positioning of Gemini 4 Argon as a general-purpose physical reasoning orchestrator across the robotics lifecycle, extending the LLM-as-controller row of the recipe-lever board to the frontier-general-model pole — Argon drives Spot (12-DoF legs + 6-DoF arm), a 6-DoF arm via code-as-policies, bimanual 76-DoF humanoids with 22-DoF five-fingered hands at 500 Hz, and the sim-data augmentation pipeline lifts real-world Franka VLA success 56→75% and completion time 35s→24s on a 77-task curriculum generated from a single RGB photo. Claims 87% on ERQA via pure reasoning. Companion to Gemini Robotics 2 brings whole body intelligence to robots with Argon above the GR2 VLA, structurally close to OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA’s zero-abstraction bet.
• [2026-10-13] InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data: InternW0-Δ (Shanghai AI Lab, Shen et al.) is the sharpest controlled instance to date of the “video co-training as supervision, not inference-time rollout” WAM recipe on the recipe-lever board — directed Mixture-of-Transformers couples a Wan2.2-TI2V-5B video expert and action expert through masked joint attention with frozen-VLM scene semantics on the action side only. Both future supervision (Causal Imprint on adjacent clean-latent differences) and geometric supervision (Track4World 4D-aware distillation) are training-only with mask-level guarantees that no activation path exposes realized future tokens to the action expert. Unified >20K-h open corpus spanning robot demos + UMI + egocentric human + Ego2Robot under a canonical state-action representation. Reports LIBERO-Plus 92.8% distribution-shift SOTA, RoboTwin 2.0 Clean2Random 71.9% / Clean2Clean 90.0%, EBench 66.0, RoboDojo 23.9% (~2× strongest prior WAM) at 152.8 ms controller-observed latency on a single RTX 5090. Same canonical checkpoint deployed on two gripper-based and two dexterous-hand platforms via embodiment-specific post-training. Contrasts cleanly with Causal World Modeling for Robot Control (LingBot-VA) LingBot-VA’s unified causal AR flow-matching and μ₀: A Scalable 3D Interaction-Trace World Model µ₀’s frozen 3D-trace WM + action-expert factorization as three structurally distinct answers to the same question of how to put future prediction into service of control.
• [2026-10-13] The Will to Act — Natura-1: Natura-1 (Nature Will) adds a unified 50B MoE for language + vision + world + action in one sequence model row to the recipe-lever board, structurally distinct from VLM-with-action-head (π*0.6, Embodied-R1.5), frozen-WFM+action-expert (µ₀, FLUX-mimic), symmetric-I/O backbone (Rho-1), and the two-head WAM+BC pattern (EgoWAM) — pretrained on 1M hours of egocentric human experience (61.7M episodes, 317K tasks, 14.3K scenes), then adapted with limited post-training to drive 20-DoF dexterous hands under a strict 100 ms control loop. First filed human-video-scale datapoint at the 1M-hour mark on the dexterous-hand target embodiment specifically (sibling to Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2 at the same scale but on YAM-arm WAM substrate; ~48× EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale on dexterous hands). Head-to-head: on dual-arm ROKAE AR5 + Wuji 20-DoF hands, neither π0.5 (fine-tuned on Nature Will dexterous data) nor GPT-6-Astra scores; on ARX AC One parallel-gripper where both baselines are fine-tuned on matched-task gripper data, π0.5 13.5 / GPT-6-Astra 16.7 vs Natura-1 Base 53.4 / Agentic 59.1 with 20-DoF dexterous hands. Scaling claim: 44K → 1M hours of egocentric data lowers action-prediction MSE on real-robot data by 67.7% without fine-tuning on robot data — first filed quantitative read on zero-shot human-video-to-robot action prediction at million-hour scale. Emergent behaviors reported in daily interactions (card-overhang graspable-edge creation, rod-as-reach-tool, mid-air wrist rotation to steady toppling stacks) are claimed as not in the prompt or in the robot demonstrations. Full-stack co-design is pitched as load-bearing: in-house motion labeling with 41% MPJPE reduction vs open-source baselines, 350K hours/week processing throughput, custom attention kernels at ~1,200 TFLOP/s (>6× FlexAttention), >50% MFU + 98% GPU utilization in production training, Rust-based inference runtime at <100 ms end-to-end latency — pushes the deployment-runtime axis (InstinctFlash — High-Performance Serving Runtime for Robotics Models InstinctFlash) from aftermarket serving into training+inference co-design. Caveat: six of the long-horizon demonstrations are flagged as using additional task-specific tuning on top of Natura-1; the remaining results share one checkpoint. Blog does not disclose expert count / activation sparsity / action head architecture / open-weights plans.
• [2026-10-12] RoboRSI: Stable, Efficient, and Reusable Robot Self-Evolution in Complex Real-World Environments: RoboRSI (Noematrix Lab + SJTU Cewu Lu group) adds a four-agent coding-harness over a four-level skill tree cell to the orchestration-over-frozen-VLA row — distinct from Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA’s single-planner-over-single-VLA shape and Addressing the Orchestration Gap in Generalist Robots via Physical Agency (Pigey) Pigey’s TAMP + VLA factorization: Manager / Planner / Engineer / Reviewer own distinct lifecycle roles over Task Family → Compound Skill → Atomic Task → Base Skill, final-feedback attribution routes along the actual call path to the earliest failing node, and revisions are confined to that node’s branch. The “policy” is synthesized code emitted by the Engineer, not a learned VLA — closer in spirit to Playful Agentic Robot Learning RATs’s Code-as-Policy skill library but with ownership-scoped TSR constraining where edits may land and a replay-window check preventing regressions on previously-working behavior. Reports SOTA on LIBERO / LIBERO-PRO / LIBERO-Plus / RoboTwin exceeding strongest baseline by 2.7–11.0 pp (abstract-level only; per-benchmark splits not retrievable at filing), real mobile-manipulator household cleanup over 104 rounds, and a stacked ablation showing Code-on vs Code-off cuts median tokens by 29.4% / VLM calls by 27.2% / wall time by 17.0% on repeated tasks (120 × 5 = 600 episodes).
• [2026-10-12] TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning: TouchScale (Li et al.) stakes a tactile-mid-training-beats-vision-only position for contact-rich manipulation: a visual-tactile backbone mid-trained on 500 h of human bimanual visual-tactile recordings (ego RGB-D + wrist RGB + 880-taxel tactile at 30 Hz) with a future-tactile-change prediction objective, then post-trained on 50 teleop demos / task, lifts real-world success on four contact-rich tasks (xArm6 + BrainCo Revo 2) from 22.5% → 57.5% average. Direct counter to the pure-vision + action-head VLA recipes that under-weight contact dynamics; sibling to Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction T-Rex’s variable-rate MoT + temporal tactile VQ-VAE on the policy side but moves the lever to the mid-training corpus.
• [2026-10-10] A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control: SGS (Zhang, Guaman Castro, Yin et al.; NVIDIA Isaac Lab + UW) distills mega-scale PPO sim-trained contact-rich manipulation policies into RGB-conditioned students that transfer zero-shot to a real UR5e on NIST assembly tasks (6/20 board assets solved), with no real-world data and no human demonstrations at any point in the pipeline. Stakes out the “pure sim-to-real distilled RGB-VLA” pole opposite the demo-centric manipulation-VLA lineage (π0.5, π0.7, GR00T N1.7, MolmoAct 2, DiT-Flow) and the retargeting-centric dexterous lineage (CHORD, DemoMimic, HOST). Also notable: the discovered strategies are visibly non-anthropomorphic (tip rod against board leg, upright gear against shaft, flip nut in place) — evidence that pure-RL student policies land on gripper-native behaviors demo-based policies by construction cannot.
• [2026-10-10] LeWAM: End-to-End World and Action Modeling with JEPAs: LeWAM (Fu/Siebert/Halicki/Balestriero, AMI Labs lineage) adds a JEPA-end-to-end-from-pixels row to the recipe-lever board distinct from VLM+action-head (π*0.6, Embodied-R1.5), unified causal AR flow-matching (LingBot-VA), frozen-WFM+action-expert (µ₀), and symmetric-I/O backbone (Rho-1): one shared ResNet-18 encoder + MoT predictor with state and action branches, trained under L_dyn + L_act + λ·SIGReg. 17M params, single GPU, serves as both direct policy and planner. Beats frozen DINOv2-CLS encoder on direct-mode control at matched tasks (86.7% vs 79.0% across four tasks) — a measured counter-example to the “freeze big foundation encoders” position, at the research-budget scale where it can be run end-to-end. On physical 7-DoF hardware across three tasks (300 trials) LeWAM runs within Diffusion Policy’s GFLOP budget (270.2 vs 3603.6 for π0.5) with competitive 33–36% avg success, trailing Diffusion Policy (38.9%) but beating π0.5 (28.9%).
• [2026-10-09] ARC: A Reasoning Recipe for Robot Foundation Models: ARC (arc-robot-reasoning.github.io) is a recipe wrapper that teaches existing robot foundation models (π0.5, Cosmos3-Nano-Policy) action-grounded reasoning with no new robot demonstrations and no foundation-scale training. Occupies a distinct cell from the companion-reasoner cluster already on the board: FailSafe polls every 10 steps on failure; Harness VLA calls an LLM planner at subtask boundaries; ARC runs an async background VLM reasoner at ~1 Hz against live camera views while the policy executes 15-action chunks at 15 Hz, and fine-tunes the policy to consume the trace (jointly updating VL backbone, trace encoder, action expert; factual traces via flow matching, contradictory traces via a counterfactual margin loss). Reports 9.5% → 91.7% real-robot success for π0.5 + ARC — sharpest filed single-recipe gain over base π0.5 on this page — plus up to +50 pp on RoboLab-Reasoning-50 and 4.3× fewer training updates to reach the instruction-only baseline on RoboLab-120. Also reports π0.5 + ARC within 0.4 pp of the ten-step base at one Euler step for action generation. Models, code, and ARC-Trace-DROID pledged after review.
• [2026-10-09] In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks: SimpleICL (Alibaba Qwen + CASIA) proposes a problem-definition-first recipe for in-context robot learning: four discrimination dimensions (action / composition / object / affordance semantics) + three robustness axes (spatial / human-detail / scene variations), paired with within-scene discrimination pairs and a 3,000-pairs/robot/day collection-group protocol, lifts RoboTwin 2.0 OOD success 40.3% → 72.1% without any model change on a Wan2.2-5B + 1B action-DiT stack. On eight unseen real-world tasks the sharp gap over Fast-WAM and π₀.₅ (0.68 vs 0.42 / 0.31) opens specifically on Hard level where multiple valid executions share one scene — the cleanest filed statement that language-conditioned VLAs cannot pick between affordance-equivalent targets without a demo. Composition emerges without explicit drill; affordance does not. Minimalist “democratized” counter to the engine-scale readings of S1: In-Context Learning for Robotics S1 and GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5.
• [2026-10-09] UniWAM: Unified World-Action Model: UniWAM (Song, Chen, Wang et al.) adds a three-head unified-backbone row to the recipe-lever board distinct from the two-head WAM+BC trunk pattern of EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data EgoWAM and the unified causal video-action stack of Causal World Modeling for Robot Control (LingBot-VA) LingBot-VA: a physical reasoner, world generator, and action predictor all first-class heads on one backbone, jointly trained from scratch on 10,000+ hours of robot + human-egocentric + VQA with a complementary-supervision pretraining recipe routing each stream to the heads it benefits most. Low-level actions represented in natural language to preserve inherited VL capability; post-training uses future visual noise augmentation (reduces reliance on precise future prediction) + history-conditioned flow matching (encoded action history initializes the sampler, cutting denoising steps). Reports SOTA across in-distribution / robustness / generalization / instruction-following / long-horizon axes (no per-benchmark numbers in abstract) and crucially a log-linear scaling law for unified human-robot co-training — the first filed quantitative characterization of how the two streams compound under a joint training objective rather than EgoScale/Dyna-2’s human-video-only scaling or Isaac 0.5’s video-vs-teleop exchange at fixed mixture. Open-sourced under Apache 2.0 with code + ModelScope weights.
• [2026-10-09] Dmytro Hrybov — OpenAI Decisions API (GPT-6 Luna) as a real-time robotics policy in MuJoCo from camera views: First filed LLM-as-controller cell using the typed-decision pole rather than a general-purpose frontier LLM — Dmytro Hrybov runs OpenAI’s Decisions API (post-trained GPT-6 Luna) as a MuJoCo real-time tabletop policy over camera views, factoring each tick into two API calls. Qualitative best run grasps and transports a cube across a divider but fails placement — reproduces the contact / precise-placement failure mode flagged by Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level and seen in Comparing Fable 5.1 to Fable 5 on robotic manipulation / Gemini Robotics-ER 2 vs Claude Opus 5 clapboard eval — ER2 threw robot arms off table, hallucinated success (Robocurve / Jay Chooi), now with a schema-constrained decoder (likely under-represents the continuous precision axis a VLA’s flow-matching head carries).
• [2026-10-08] RoboJEPA: Scaling Robotic Latent World Models: RoboJEPA is deployed zero-shot as a robotic agent via goal-image planning on real hardware for long-horizon tasks, from an 8B JEPA world model pretrained on 12 robot embodiments. Occupies the “WFM-as-planner-directly” deployment pattern — the world model itself is the policy via goal-conditioned latent planning — alongside the DVA “video model as policy” line (Rhoda AI — Web-video pretraining scaling laws for robot policy learning (DVA), Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners) and distinct from the WFM-as-reward / WFM-as-policy-internal-substrate patterns tracked under World Foundation Models.
• [2026-10-08] Reality Check — Pokeandwiggle's Real-World Robot Manipulation Leaderboard on an FR3 Duo Farm: Reality Check (Poke & Wiggle) is a new centralized real-hardware third-party eval surface that puts π0.5, MolmoAct 2, GR00T N1.7, and DiT-Flow on the same bimanual Franka FR3 Duo farm under matched fine-tuning from each developer’s public checkpoint (DiT-Flow from an ImageNet-only vision encoder, having no public default), with physics-verified per-rollout object placement, nested D10/D100/D300 data tiers, and nominal/interpolation spatial generalization. The design choice relevant to the recipe-lever board is strict Excellent-only teleop training data (no recoveries, no suboptimal episodes) — aligns with Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models Spirit-v1.5’s “clean teleop is the lever” position and against the on-the-robot failure-data substrate position of RW-RL Dataset: Real-World Reinforcement Learning for Robots / OopsieData — a multi-lab dataset of real robot manipulation failures OopsieData. Isolates data-quantity scaling under held-constant data-quality cleanly from “robust to messy demos” axis, complementing the single-task fine-tuning sweep in Empirical results fine-tuning π0.5 on a real manufacturing task. Leaderboard cells render client-side so no model-ranking numbers readable at filing; structural release numbers (14,400 rollouts, 30 per combination, 95% CIs, double-reviewed annotation) establish the methodology this concept has been missing a shared-protocol reference for.
• [2026-10-08] Long-WAM: Scaling the Context of World-Action Models: Long-WAM is a WAM-flavored VLA that pairs causal future-video latent prediction with action-chunk emission under a hard real-time budget (107.4 ms/chunk on RTX 5090 incl. future-video prediction). Load-bearing claim for the VLA side: long visual context delivers large gains (RoboCasa GR-1: 63.3→78.7% at 19.2s context) only when the backbone was autoregressively pretrained, not bidirectionally. Direct head-to-head: 95% on dynamic cup stacking vs. 0/20 for both π0.5 and Fast-WAM: Do World Action Models Need Test-time Future Imagination?, contradicting Fast-WAM’s “no need to imagine future video at inference” conclusion for reactive long-horizon tasks.
• [2026-10-07] FLASH: Fast Learning via GPU-Accelerated Simulation for High-Fidelity Deformable Manipulation in Minutes: FLASH ships a sim-only distilled vision+proprioception policy for cloth folding that reaches 85.8% (91/106) real-world success on AdamU dual-arm towel folding with zero real demonstrations and 50–600 min of training wall-clock per task on a single RTX 5090 — a sharp counter-recipe to the end-to-end pretrained-VLA framing that dominates recent filings. The pipeline is teacher-student distillation (privileged-state heuristic teachers → vision+proprioception student with physics and visual domain randomization) rather than a scaled pretrained foundation model. Companion to Enabling Robust Cloth Manipulation via Inference-Time Simulator-in-the-Loop Refinement SILR, which wraps a similar base policy with inference-time MPPI refinement over the same FLASH backend to push single-arm diagonal folding from 1/10 → 9/10 on real hardware.
• [2026-10-07] World Action Learning via Interaction-Centric Spectral Latent Guidance: WING is a VLA whose action generator attends to frequency-domain guidance tokens (lowest-K DCT coefficients of a learned interaction-centric latent-action sequence) via gated residual updates, with the guidance target predicted at inference from current image + language + robot state alone. Reports 99.20% LIBERO, 93.80% RoboTwin 2.0, 57.7% RoboCasa-GR1 average success. Adds a spectral-conditioning row to the recipe table that otherwise contains in-context prompts (DM0.5, Zero-WAM), reduced action spaces (Translation-as-Bridging-Action), latent foresight (InternVLA-A1.5), and object/intent conditioning (LUCID).
• [2026-10-06] OCBench — A Controllable Robotic Manipulation Benchmark for BC and RL: OCBench is the shipped benchmark/code artifact for the Behavioral Cloning Mystery line — gives the VLA action-head design axes (chunking, history conditioning, model size, feature scaling) a controlled sim substrate where fresh non-Markovian human-like demonstrations can be streamed under MJWarp.
• [2026-10-06] RealtimeWAM: One-Step Asynchronous World Action Models: RealtimeWAM (Lv, Du, Feng, Yong, Gu, Yang, Gong, Ren, Zhang, Wang; NTU + SenseTime) extends the async-inference axis from between policy chunks (πR²: Reactive Real-time Flow Policies πR², Real-Time Robot Execution with Masked Action Chunking REMAC) to between the two experts within a single WAM forward pass via Cross-Expert Wavefront Pipelining — block-wise sharing of the video KV cache with synchronization deferred until the action expert’s attention consumes each block. Stacked with Teacher-Anchored Consistency Distillation (one-step action student anchored against a frozen multi-step teacher’s rollout endpoint to close the local–global error gap), reports ~25× end-to-end H100 speedup at <1% drop across LIBERO, LIBERO-Plus, and RoboTwin on both Fast-WAM and Faster-WAM. Finer-grained parallelism axis than any filed async-VLA recipe so far; applied to the MoT WAM substrate rather than to a VLA with a single action head.
• [2026-10-06] Arm-wise Compositional Generalization in Dual-Arm Vision-Language-Action Models: AE-VLA (Zhang et al., Evolving Machines Lab + collaborators) adds a shared-policy architectural-structure cell to the recipe-lever board specifically for dual-arm compositional generalization — a new axis none of the existing entries targets. ACG-Bench supplies 23 task-condition pairs (6 in-domain + 17 held-out across reorder / sync / sync+reorder / cross-task) with every method receiving the same per-arm atomic prompts from a shared rule-based phase scheduler, isolating execution of a supplied composition from plan inference. On a shared π₀.₅ backbone, AE-VLA (arm-token grouping + SkillLoRA bank-of-10 skill-specific LoRA adapters with per-arm routers + arm-wise attention that blocks direct cross-arm action attention while preserving shared global context) reports 21.53% simulation generalization success vs 2.94% Single π₀.₅ / 3.06% MA-VLA / 5.53% two-independent-π₀.₅ — the Dual π₀.₅ baseline is the sharpest filed challenge yet to the “give each arm its own policy” counter-recipe (independent per-arm policies still fail at coordination because local obstruction and goal-fragment success don’t compose). Real SO-101 across five unseen conditions: 39% mean vs 10% strongest baseline; 12/20 Stack Bowls / Sync vs 0/20 for all three baselines. Known trade-offs: pure sync remains hardest (8.67% sim), in-domain success drops vs Dual π₀.₅ (27.17% vs 41.33% sim), and on real Push Cubes / Sync Dual π₀.₅ wins 6/20 vs 2/20. Three design choices compose super-linearly (grouping-only 3.82% → grouping+SkillLoRA 6.00% → grouping+AWA 5.82% → all three 21.53%). Code Apache-2.0.
• [2026-10-06] OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies: OpenRUA (Chu, Barr, Le Goues, O’Hearn, Harman, Sarro, Ye; UCL + CMU) stakes out the sharpest zero-abstraction position on the LLM-as-controller row — no bespoke harness, no primitive library, no VLA, just terminal access to ROS 2’s native CLI + client library. Claude Code + Claude Opus 5 reports 99.0% on CaP-Bench and 87.0% on LIBERO-PRO. Direct counter to CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation CaP-X’s “performance rises with primitive abstraction” takeaway (headline comes from removing scaffolding instead of adding it) and sibling-but-opposite to RobotUse: Allocating Computation, Context, and Decisions RobotUse filed in the same Slack message (which bets on more harness structure: Main/Subagent/Backend + persistent playbook). Emergent-behavior rates diagnose where the gains come from: 96.80% raw-sensor processing programs, 95.87% motion-control clients, 50.13% closed-loop control programs — none handed to the agent as primitives. Connects the LLM-as-controller thread directly to Agentic Software Engineering substrate.
• [2026-10-06] RobotUse: Allocating Computation, Context, and Decisions: RobotUse (SNU, Kwak) adds a three-role factorization (Main agent / Subagent / Backend) to the LLM-as-controller row — distinct from AGP’s program-writing-over-EE-pose-interface and Show-Harness’s semantic-action-unit vocabulary. The interface bet is “visually select target and gripper pose, Backend owns geometry + motion planning + control”; Subagents retain detailed per-subgoal interaction and return compact summaries to keep the Main agent’s context small across long horizons. First filed harness-level number on the same RoboLab benchmark Sid runs VLA checkpoints on: 45% task success, +6.7pp over CaP-X.
• [2026-10-05] Rho-1: Collapsing the Multimodal Stack: Rho-1 (Reka Labs) stakes out an end-to-end symmetric-I/O backbone with actions as a native continuous modality row distinct from both the VLM-with-action-head pole (π*0.6, Embodied-R1.5-VLA, Spirit-v1.5) and the frozen-WFM-plus-small-action-expert pole (µ₀, FLUX-mimic, World Pilot): one 19B MoT jointly trained under NTP + flow-matching so physical actions and future video frames decode from the same latent state. Textual chain-of-thought and action channels share the same KV cache, so there is no separate planner/policy split. Direct scaled instance of the Sitzmann “generative-rollout pre-training is the right substrate” position (The flavor of the bitter lesson for computer vision). No LIBERO / RoboLab / RoboArena numbers; qualitative contact-aware grasping demos (gripper closes on and preserves rigid-body objects without interpenetration under text-only conditioning) are the only reported evidence. Pairs with Reka IDM (Reka Inverse Dynamics Model for Interactive World Models) for web-video → action-token pretraining substrate.
• [2026-10-03] EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras: EyeRobot 2.0 (Berkeley; Kerr/Hari/Kanazawa/Goldberg + Malik + Karen Liu) adds a gaze-as-first-class-action row to the recipe-lever board distinct from the viewpoint-decoupling neighbors LIME: Learning Intent-aware Camera Motion from Egocentric Video LIME (viewpoint SE(3) target head one-shot prior to a frozen VLA) and ActiveMimic: Egocentric Video Pretraining with Active Perception ActiveMimic (joint pretraining of viewpoint + wrist actions from body-worn RGB): a two-stage RL pipeline (per-task gaze controller trained with PPO against detected 3D object positions, then a target-selector RL agent co-trained with the BC gripper policy via EyeRobot 1.0’s BC-RL loop) runs on standard stereo hardware with no gaze demonstrations and no mechanical eye. On seven real-world bimanual YAM tasks with over 1,000 rollouts, No-Gaze stereo 27% → Ego + Wrist 52% → EyeRobot 2.0 67%; matches wrist cameras on unobstructed tasks (marker, tea, toaster, wrench) and beats them on tool-occlusion tasks (pot lid, boba straw) where the grasped object blocks the wrist view. The load-bearing architectural claim is that expressing both observations and actions in the moving gaze frame is the lever: a sim ablation that swaps fixation-relative actions for world-frame SE(3) collapses performance to roughly the no-fixation policy — the paper’s interpretation is that without fixation-frame canonicalization the model learns proprioceptive shortcuts without actually using the images. Matching the fovea’s effective resolution with uniform tokens would require >10× the token budget; marker (acuity-bottlenecked) sits near 0% success without fixation. Distractor-condition evaluations under identical initial object positions show EyeRobot 2.0 reaches for the correct object when gaze is directed at it and mostly misses grasps when gaze is wrong, while wrist cameras commit to the largest/most-distinct in-view object and cannot recover — the sharpest filed architectural answer to the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position on the “decouple vision from action” axis rather than the modality-addition axis. MuJoCo YAM twin built for high-volume ablations under matched hyperparameters/camera/robot config (not for sim-to-real), full sim setup pledged open-source.
• [2026-10-01] Argus: An Open-Source Annotator for Robotics Data: Argus (Pantheon Phase 2) extends the WAM-vs-VLA data-cleaning debate by making the recommendation explicitly objective-dependent: operator-mistake episodes should be filtered or trimmed for imitation learning (behavior cloning will copy the suboptimal approach) but kept and labelled as recovery data for world models. First filed articulation that “data quality” splits by downstream training objective rather than being a single-threshold decision. Augments the earlier predecessor’s WAM-first severity framing with a VLM-driven detection method for the context-dependent errors (meandering, label mismatch, undone goals) that deterministic checks miss.
• [2026-10-01] Eren Chen teases Physical Intelligence π0.8 announcement: Eren Chen (Physical Intelligence) teases a forthcoming π0.8 release — bare tweet + screenshot, no linked artifact at filing. Suggests the π-line cadence (π0 → π0.5 → π*0.6 → π0.7 → π0.8) is continuing through late 2026; flagged for /bud refresh once PI publishes the companion blog / tech report so this concept page can place π0.8 on its recipe-lever board.
• [2026-09-30] SIMPACT: Simulation-Enabled Action Planning using Vision-Language Models: SIMPACT (Liu, Yao, Chen, Gao, Mao, Huang, Du) reports the sharpest π₀.₃ negative result on file: 0% success across five contact-rich real-world tasks (non-toppling push, bowl stacking, pivoting, shape rope, shape Play-Doh) — the paper argues VLA generalization “poorly to out-of-domain, challenging tasks” and its zero-shot 60–90% counter-recipe is a VLM + externally-constructed multi-physics simulator + in-context planning loop with no VLA in the stack. Sibling to Guava: An Effective and Universal Harness for Embodied Manipulation Guava from the same first author on the “no VLA at all” cell but with the physics understanding living in an external simulator queried at test time rather than in a small distilled model’s weights. Direct empirical support for the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position that VLA scaling has not solved contact-establishment tasks — and a concrete measurement that swapping to physics-in-the-loop closes the gap on the same hardware.
• [2026-09-30] Guava: An Effective and Universal Harness for Embodied Manipulation: Guava (Liu et al., UMD/UIUC/Waterloo/MBZUAI/UPenn/Amazon FAR) stakes out the “no VLA, VLM + semantic tools + distilled compact model” cell of the recipe-lever board — a Qwen3.5-4B post-trained via SFT + GRPO on <2K frontier-VLM ReAct trajectories (GPT-5.4 under the Guava harness) reaches 75.6% overall on 15 sim tasks, beating its own teacher (70.2%), and transfers zero-shot to real Franka at 86%/92% ID/OOD. Direct parallel to Show-Harness: Just a VLM Agent Can Play Robots Show-Harness’s Qwen3.5-2B distillation result but with an explicit long-horizon RL stage as the mechanism for the biggest gains (shell game 6.7→60.0%, place all red objects 0.0→93.3% via GRPO with sparse task-success reward). Load-bearing new datapoint on the “explicit recovery trajectories via perturbed-state augmentation” data axis — recovery behaviors emerge for unseen real-world joint-limit and unreachable-pose failures, evidence the RL-trained policy learned to reason over execution feedback rather than memorize correction patterns.
• [2026-09-29] DexAgent: An Agentic Human2Sim2Robot Framework for Dexterous Manipulation with Self-Evolving Tool Library: DexAgent (Stanford; Wang, Li, Fei-Fei, Wu, Huang) is a data-side contribution rather than a policy-architecture one: an agent-authored Human2Sim2Robot pipeline that produces dexterous-manipulation training data from a single egocentric human video, with property-specific tool selection + verifier gates, and reports 3.5× real-world success over baselines across 11 tasks when policies are trained on the resulting data. Complements the pretraining-substrate work on this page (ActiveMimic: Egocentric Video Pretraining with Active Perception, ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation) with a per-video agentic data-generation option scoped to the dexterous regime.
• [2026-09-28] Delta Intelligence launches Δ₀ — whole-body humanoid model with latent world-action brain, 69-DoF controller, and 10,000+ h egocentric panoramic pretraining: Delta Intelligence’s Δ₀ launch stakes out a whole-body humanoid + latent-world-action-brain + learned 69-DoF-controller cell of the recipe-lever board no filed system currently occupies at this scale — closest neighbors are Gemini Robotics 2 brings whole body intelligence to robots Gemini Robotics 2 (closed, no scaling numbers), HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own HomeBody (whole-body G1 via LLM-as-controller with five typed skills, not a learned 69-DoF controller), and Isaac 0.5: An Open-Weight Embodied Foundation Model Isaac 0.5 (clean scaling, bimanual not whole-body). The 86% sim vs 89% real dishwasher parity at matched data scale is the sharpest sim-to-real fidelity number on this page and directly answers whether R2S2R actually works for policy scaling (Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies). The 20% → 65% real-world RL lift with human-in-the-loop corrections scales the MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory / Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact / Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics intervention-scaling thesis to a whole-body humanoid. Delta D1 head-mounted panoramic capture device pretrains on 10,000+ h with a shared cross-embodiment action space — the “make the collection tool embodiment-invariant” position (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic, HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet) HandUMI, Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1, TwinDEX: A Scalable Path from Robot-Free Data Collection to Real-World Dexterous Manipulation TwinDEX) pushed to whole-body humanoid pretraining. Launch-thread only — no head-to-head, backbone size, action-head architecture, or open weights.
• [2026-09-28] We Looked at the Data — Diagnosing Open-Source Robotics Datasets for WAM Training: Pantheon’s dataset audit takes an explicit WAM-first vs VLA-agnostic severity framing that sharpens the “how much does data cleaning matter” debate on the recipe-lever board — direct counter-position to Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models Spirit AI’s “clean data is the enemy of great robot foundation models” thesis (which was aimed at VLAs) by arguing that for WAMs specifically many anomalies directly poison the action→observation causal signal and must be filtered/masked. Per-episode manifests released for 4 open datasets (FastUMI_100k, MolmoAct2-BimanualYAM, 10Kh-RealOmin, XDOF/ABC-130k) that overlap most VLA/WAM pretraining substrates on file, making the manifests an immediate upstream filter for anyone training on these corpora.
• [2026-09-28] EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics: EmbodiedSWE claims a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on a real robot, with generalization to held-out task variations improving as agent-authored diversity increases. Sharpens the “where does VLA training data come from” question — this is the first filed VLA that trains entirely on coding-agent-authored simulation trajectories (via EmbodiedSWE-Gen) rather than teleop or motion-planner scripting, with the coding-agent-as-data-engine acting as an alternative to VLA-native RL for long-horizon dexterous tasks.
• [2026-09-28] HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own: HomeBody (Stanford Movement Lab + Caltech) stakes out the whole-body-humanoid + persistent-memory cell of the LLM-as-controller row on the recipe-lever board — GPT Astra directly calls a fixed library of five typed motor skills (navigate / pick / place / open_drawer / pick_from_drawer) on a real Unitree G1 across an unseen kitchen, with the same VLM constructing an Isaac Sim digital twin from the robot’s own exploration data as external spatial memory. Structurally between Robot-Use Agents Isola’s positional “robot-use agents” essay and its scaled-up empirical demonstration — the paradigm on a full humanoid with load-bearing spatial memory over an unseen multi-room space. Consistent with Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level Malik’s sensorimotor counter-position on failure modes: drawer opening + finger-servo overheating are exactly the contact-rich regime that predicts LLM controllers to struggle. Qualitative demo only — no success rates, latency numbers, or head-to-head vs a trained VLA on the same G1.
• [2026-09-25] Neural Scaling Laws in Robotics: Sartor & Thompson meta-analysis (n=327 papers, VLA-dominated at n=93 data-scaling studies) quantifies the absence of controlled model-size scaling in the VLA literature: only 22 papers of 327 report model-size scaling at all, median within-study parameter sweep 10.3× (a single order of magnitude), fitted average $\alpha_N = -0.246$ pooled from 34 heterogeneous studies. Confirms yixuan.ren’s Slack observation that VLA-scaling papers focus overwhelmingly on data-axis sweeps with side-experiment model-size ablations, and that no filed VLA paper as of mid-2024 has produced the controlled joint $(N, D, C)$ isoFLOP fit routine in LLM/diffusion research. Mean top task-success across 424 scaling studies is 67% (SD 24%), far below the 99.X% industrial-deployment target — meta-level evidence that current VLA recipes (all levers on the page) sit deep in the diminishing-returns tail of every measured scaling axis.
• [2026-09-23] Empirical results fine-tuning π0.5 on a real manufacturing task: Dream Machines publishes the sharpest single-task empirical fine-tuning map of π0.5 to date — 11 controlled experiments on a 21h bimanual pick-and-insert dataset (n=40 rollouts each with binomial CIs), all datasets released on HuggingFace. Baseline full FT at 76%; LoRA at rank 64/256 underperforms by 23–43 pp even under Schulman’s all-linear + 10× LR-multiplier recipe (counter-datapoint to the LoRA Without Regret “LoRA matches FullFT” claim, author-hypothesized as “new embodiment is not a low-rank edit”); scene diversity across five collection scenes beats homogeneous evaluation-scene data by 30 pp at matched 4h; one clean hour lifts the 21h baseline from 76→90%; 240 DAgger-style intervention rollouts on a 1h seed policy reach 88% real-eval and beat the 21h full-data model by 12 pp (extends the Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact / MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast) MicroFactory / Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble) Robo Robotics intervention-scaling thesis with per-experiment sample-count data); tuning only RTC serving parameters (reinfer=50, blend=20) on a fixed checkpoint lifts 76→93% and 90→98% (test-time counterpart to the training-side async-fix line of Real-Time Robot Execution with Masked Action Chunking REMAC / πR²: Reactive Real-time Flow Policies πR²); and training from PaliGemma-init-only (no π0.5 weights) yields plausible motions but 0/40 successful insertions — sharpest published rebuttal of the original π0 paper’s claim that pretraining is only marginally load-bearing.
• [2026-09-23] TacBPM: A Tactile-conditioned Behavior Prior Model for Dexterous Reorientation: TacBPM parallels the frozen-base residual-adaptation thread (Proxy Policy Steering: Adapting Frozen Robotic Foundation Models via Velocity-Space Proxy Residuals) in the dexterous-manipulation RL regime rather than the VLM+action-head regime: distill sphere-manipulation specialists into a tactile-conditioned behavior prior, freeze it, and let downstream policies emit residual latent actions on top rather than raw joint commands. Not a VLA in the classical VLM-backbone sense, but relevant as a recipe-lever data point — the “frozen prior + residual latent head” architecture generalizes beyond visual-language conditioning to tactile-proprioceptive conditioning, and the training-acceleration + raw-PPO-fails-here claim is the same shape as the base-model-priors-are-broad-and-useful argument PPS makes on the training side.
• [2026-09-23] InstinctFlash — High-Performance Serving Runtime for Robotics Models: InstinctFlash (General Instinct, YC P26, AGPL-3.0) is the third VLA-runtime release on this page — spans eight model families (π0.5, LingBot-VLA-4B, LingBot-VLA-V2-6B, LingBot-VA WAM, GR00T-N1.7-3B, Cosmos3 Edge/Nano DROID, DreamZero DROID) under one Runtime API, with a planner that reads each checkpoint’s declaration and decides which passes are provably valid without loading weights. Contrasts with the two priors: Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots Embodied.cpp maximizes portability (Apache-2.0 llama.cpp-shaped C++ + GGUF, CPU/CUDA/NPU), APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin APXinf-robo maximizes per-family Jetson performance (Rust + hardware-locked kernels, one family well). Reported Jetson Thor P50 speedups vs PyTorch: 7.88× π0.5 FP8, 5.36× LingBot-VA FP8 25V/50A, 33.78× LingBot-VA composed with 2V/4A distillation. GR00T N1.7 comes in at 1.19× under strict BITEXACT — a low-headroom outlier arguing not every VLA has room to move under strict precision. Companion launch reports GI0 (their own diffusion-transformer VLA conditioning actions on predicted future videos) at 5.75× wall-clock advantage over GPT-6 Astra on the same task, one side-by-side. All three VLA runtimes now converge on the OpenPI wire protocol — unmodified openpi-client connects unchanged — the first infrastructure standardization on the deployment-runtime axis of the recipe-lever board. No closed-loop task-success numbers reported at launch (unlike APXinf-robo’s 500-episode LIBERO-10 evaluation), quality deferred to the validate non-inferiority certificate.
• [2026-09-22] Grounded Action Model: 3D Grounding as a Foundation for Robotics: GAM (Zhang, Huang, Shailesh, Peng, Duan, Krishna) adds a 3D-grounding-first row to the recipe-lever board structurally distinct from action-pretraining (π*0.6), unified-VLM (Embodied-R1.5), WAM+action-expert (µ₀), and general-video-pretraining (Isaac 0.5): a pretrained 3D grounding model (WildDet3D; SAM3D as plug-in) receives language/point/box prompts, produces object-centric image tokens + detection tokens carrying selected-object point clouds and metric geometry, and a multi-stream MM-DiT flow-matches action chunks on top with robot state history. RoboTwin 2.0 (50 tasks) 55.3% avg vs 52.0% Spatial Forcing and 47.6% under randomization vs 30.4% Abot-M0 — action policy trained only on clean-scene demos. LIBERO-PRO (16 perturbation settings) 61% avg vs π₀.₅ 53%; LIBERO-Spatial with target change 88% vs π₀.₅ 1%. Bimanual YAM under visual shift 17/20 vs π₀.₅ 4/20. GAM + Molmo2 planner on Franka 64.7% ID / 49.8% OOD step completion on long-horizon and memory-dependent tasks. Sibling to LIT — two-stage VLA/WAM training constrains visual conditioning via latent interface to prevent visual shortcuts LIT (same lead author, Duan) — LIT bottlenecks vision through a learned latent interface recovering a spatial goal; GAM bottlenecks it through a pretrained 3D grounding module whose output is metric geometry. Both attack the same “action expert picks up spurious visual correlations” problem from different layers of the stack. Author-flagged limitation: grounding errors propagate; unselected obstacles are omitted context.
• [2026-09-21] SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation: SkipVLA (Agrawal, Rahman, Yeh, Kingston; Purdue) adds a hybrid-with-classical-motion-planner row to the recipe-lever board distinct from the VLA-only levers already on the page and from the Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA “frozen VLA as narrow retryable primitive under an LLM planner” recipe: segment the task so a classical collision-free motion planner handles free-space transit and the frozen VLA is called only for contact-rich skills (grasping, placing). Glue is a lightweight target-pose predictor that reuses the frozen VLA’s own vision-language backbone and is trained without additional demonstrations by distilling what the underlying VLA already knows. Reports up to 2.5× faster task completion at matched task success across three VLAs on 13 LIBERO tasks, and matched real-world success with significantly lower energy consumption on three pick-and-place tasks on a 6-DoF YAM arm. Direct instance of the “reduce VLA call count via task-structure segmentation” answer to the VLA-latency question, complementary to the per-call speedup axes (APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin quantized serving, C³ache: Accelerating World Action Models with Cross Inference Chunk Cache cross-chunk caching, TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM LLM-free architectures) and the async-chunking axes (πR²: Reactive Real-time Flow Policies πR², Real-Time Robot Execution with Masked Action Chunking REMAC).
• [2026-09-20] CausalWM: Causal Chain-of-Thought Reasoning for Embodied World Model: CausalWM (Aether Labs) is the second filed instance of the modality-autoregressive WAM recipe first proposed by ModAR — 16B shared diffusion Transformer chaining optical flow → 3D pointmaps → future RGB with a causal-across-modalities attention mask. Language instructions and robot actions (rendered as URDF control videos) are complementary conditioning channels for synchronized head + wrist views. First place on TriWorldBench across 36 published models; SOTA on PAI-Bench robot-domain language-conditioned generation.
• [2026-09-18] Action Chunks Skrrt Skrrt: Second Binh Pham garden.binhph.am article (after Closed-Loop Policies), per Sid’s Slack framing a “timely post on speeding up robot movement during inference” — sits on the async-inference / chunk-latency axis alongside REMAC / πR² / A2A / Legato on the training-time side and Embodied.cpp / APXinf-robo / C³ache / TurboVLA on the runtime side. Article body was unreachable at filing (HTTP 429 across ~10 retries); page filed from Slack summary + author context. Which specific latency lever the article uses is not recoverable at filing time — flagged for /bud refresh once garden.binhph.am is reachable.
• [2026-09-18] APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin: APXinf-robo (RLinf / Infinigence; same lab as Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA) is the sibling VLA-runtime release to Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots Embodied.cpp on the substrate axis — Rust + CUTLASS + FlashAttention with hardware-locked kernels and Jetson-family FP8/INT8 specialization vs Embodied.cpp’s llama.cpp-shaped portable C++ + GGUF answer. Ships π0.5 first with 41.16 ms / 24.3 Hz Jetson Thor FP8 (vs 72.45 ms / 13.8 Hz BF16, 1.76× speedup), 26.32 ms / 38.0 Hz with onestep action-generation pruning, and 92.0–92.8% LIBERO-10 across three precisions vs the pi05_libero_base 92.4% reference (500 episodes, published protocol) — quantized VLA serving on Jetson preserves policy quality within noise. OpenPI-compatible websocket protocol makes it a drop-in replacement for the reference π0.5 serving stack, and per-call latency of 26–41 ms makes the RLinf lab’s own harness-style “frozen π0.5 as callable primitive” thesis economically plausible in a way a 400 ms PyTorch inference does not. Ships an agent-driven model-port-workflow skill (Claude Code / Codex) with GR00T N1.7 as the reference porting example — early instance of “agentically optimized CUDA kernels” as a shipping deliverable.
• [2026-09-18] PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics: PointZero is a data-efficient VLA-adjacent recipe on the low-supervision end: an action-free 3D-track pretrained trunk (2.9M synthetic frames, no robot actions) plus a lightweight action head trained with only 20 action-labeled demos + 100 actionless demonstration videos per task. Beats DP3 on all 7 evaluated sim + real manipulation tasks and matches or exceeds 3PoinTr on 6/7 (Blockstack sim 99.8 vs DP3 44.4 / ATM 40.0 / 3PoinTr 90.9; Drawer real 100 vs 90 / 30 / 70). Ablation isolates the pretraining objective as load-bearing: scratch vs pretrained at matched 20-demo budget improves Glass from 66.6→79.5% (all-parameter FT) and 49.1→64.6% (frozen point-processing stream) — the gain is largest on the harder tasks and when the pretrained stream is protected from IL gradient noise. Sibling to the action-free-pretrain family (V-JEPA 2-AC, LingBot-VA’s latent-action tokenizer) but with 3D point tracks — not video latents or pixel futures — as the pretraining signal.
• [2026-09-18] ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning: ORCA (Christoph et al., ETH Zurich; IROS 2025) provides an open-source 17-DoF anthropomorphic hand as a candidate VLA target embodiment, validated with a diffusion-transformer imitation policy running 7h 17min of continuous autonomous pick-and-place with no hardware intervention (~2,000 grasp cycles) and a zero-shot sim-to-real A2C policy for in-hand ball reorientation. Complements Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand Aero Hand Open as an open-hardware target for open VLA policy deployment, at the higher-DoF / dedicated-tactile end of the substrate spectrum.
• [2026-09-17] Agent as Policy for Robotic Manipulation: AGP (Jia, Lin, Zhang, Zhang, Liu, Jiang; Notre Dame) is the paper-scale version of the LLM-as-controller row on the recipe-lever board — same substrate as Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness and Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK (GPT-6 Astra + EE-pose + tool harness on YAM arms), now with a controlled 8-task-configuration evaluation, 162 released real-robot trials (10 tasks × 6 models × 3 reasoning-effort levels on HuggingFace), and quantitative time/complexity accounting. Reports ≥80% success in 7/8 configurations (100% in 5), including 100/100/80% block construction and 0.73 mm-clearance assembly from human-video specification. Sibling to Show-Harness: Just a VLM Agent Can Play Robots on the same row but with a distinct interface bet — program-writing over a typed EE-pose/joints/gripper tool interface rather than semantic action-unit vocabulary. Deformable-object simultaneous-arm towel folding (3/5) reproduces the contact-establishment failure mode Comparing Fable 5.1 to Fable 5 on robotic manipulation flagged on the same hardware family, consistent with Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level‘s sensorimotor counter-position for this class of tasks. Task-experience reuse (saved procedures + programs shorten repeated trials) transplants the experience-library-reduces-required-interaction-depth finding from text tool-use to physical execution.
• [2026-09-16] Modality-Autoregressive World-Action Models: ModAR (Hung, Duisterhof, Ramanan, Ichnowski; CMU) adds the modality-autoregressive row to the multi-stream WAM board, distinct from Flex-π: A Multi-Stream World-Action Model with Compute Flexibility Flex-π’s shared-latent parallel denoising and EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data EgoWAM’s single-target ablations: denoise depth → DINO → point tracks (→ optional RGB) as an autoregressive chain where each modality conditions on previously generated ones. Empirically: point tracks + DINO + depth are load-bearing, adding RGB does not consistently help, and modality-AR beats parallel-denoising WAMs at every data scale evaluated. Trained-from-scratch ModAR matches fine-tuned Flex-π (75% vs 72% observed avg) at ~20× fewer FLOPs on the same data — a small-scale datapoint for Going Beyond World Models & VLAs Florence’s “video pretraining is a crutch” position. Three-real-bimanual-task evaluation with human-video co-training improving further, consistent with EgoWAM’s finding that the WAM head is the channel through which human data enters the policy.
• [2026-09-16] Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models: LIT arxiv paper (Lin, Shailesh, Luo, Duan; 2609.12641) is the peer-reviewable form of the LIT — two-stage VLA/WAM training constrains visual conditioning via latent interface to prevent visual shortcuts announcement thread already on file, confirming the load-bearing method: Stage 1 trains the action expert image-free on language + robot state + terminal SE(3) end-effector pose per chunk; Stage 2 introduces a latent interface that aggregates visual + semantic representations as the only visual pathway to the action expert, supervised to reconstruct the Stage 1 terminal pose. Framework-agnostic across four backbones (π0.5, MolmoAct2, FAST-WAM, ImageWAM); LIBERO-Plus +3.87–10.70 pp with LIBERO preserved/improved; real-world +13.30–16.70 pp aggregated across unseen camera configs, lighting variations, and distractors on three tasks.
• [2026-09-14] RewardAI OM-1 + Omnibody Hand launch — zero-shot cross-embodiment dexterous manipulation from human data (DexCap successor): RewardAI’s OM-1 launch stakes out a dexterous-hand-hardware + human-mocap-data row on the recipe-lever board: 7-DoF Omnibody Hand as the fixed target embodiment paired with a DexCap-lineage glove-based human demonstration capture pipeline, and one policy zero-shot generalizing across small arms, industrial arms, and humanoids without per-embodiment finetuning. Sibling in framing to Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic (M1 hand + U1 mechanically-locked wearable) and Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1 (100K-hour UMI parallel-gripper corpus) but distinct in keeping the demonstrator’s bare hands + gloves and pushing invariance onto the 7-DoF hand+policy pairing. Launch-thread only; no per-task success rates, backbone/action-head architecture, teleop-hour count, or head-to-head with π0.5 / GR00T / LingBot / RealDex1 disclosed. Real-time execution claimed on quadmanual phone packaging, bartending, laundry folding, Ethernet-latch unplugging, and humanoid whole-body manipulation.
• [2026-09-14] HuRo: Robotizing Human Videos for Scalable VLA Pretraining: HuRo (Yonsei; Jeong et al.) adds a robotized-human-video-as-primary-pretraining-substrate row to the recipe-lever board: a single pipeline visually transforms heterogeneous human videos into robot-style observations and retargets hand motion into robot actions, aggregating five sources into 630K episodes / 142M frames — then pretrains an end-to-end VLA on the robotized stream. Reports 51.5% → 80.3% overall completion and 34.9% → 72.2% OOD completion under spatial/visual shifts across four real-world manipulation tasks as robotized data scales, with the OOD gain (+37.3 pp) exceeding the in-distribution gain (+28.8 pp). Load-bearing ablations: visual robotization specifically improves OOD robustness and end-to-end pretraining with retargeted actions beats visual-only transfer — direct empirical answer to whether the action-retargeting channel adds signal beyond observation alignment. Sibling to Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data Ego2Robot on the “render human video into robot-format training data” branch, distinct from ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining‘s absorb-the-gap-in-the-training-objective answer.
• [2026-09-12] 2AM: Grounding Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation: 2AM (Hu, Chen, Cao, Detry) adds an interface-bandwidth-as-the-lever row to the wrap-frozen-VLA-with-orchestrator cluster distinct from Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA (staging discipline over narrow contact-rich primitives), EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents EmbodiedSkills (shared executable-skill interface with prereq + verification), and Learning What to Say to Your VLA: Mostly Harmless VLA Steering LFP (test-time language-rewrite steering): a multimodal Agent is the sole holder of task memory and compiles history into subtask language + optional 2D grasp/place/move hints at different time scales; the Action Model is a single RGB-based, episodically stateless VLA trained on hint-augmented demos under condition dropout, spatial noise, and temporal jitter to tolerate imperfect Agent outputs. Reports 76.3% average completion on LIBERO-Mem (+61.5 pp over the strongest reported baseline of 14.8%), 63.0% relaxed and 11.8% strict — the sharpest filed datapoint yet that on memory-dependent long-horizon tasks the missing lever is Agent→VLA interface bandwidth, not tool library breadth (EmbodiedSkills itself reported only 12.5% on four RMBench memory tasks with the same orchestrator-loop pattern). Deliberately drops depth, online geometry, and planner-based object motion to isolate the interface variable.
• [2026-09-11] Proxy Policy Steering: Adapting Frozen Robotic Foundation Models via Velocity-Space Proxy Residuals: Proxy Policy Steering (anonymous under review; @tritiumac tweet) stakes out a training-side frozen-base adaptation recipe distinct from LoRA / RECAP / RL²-VLA / LFP / VINE / QGF: train an on-policy-distilled reference proxy v_ref that reproduces the frozen base’s denoising behavior on the deployment support, initialize a task proxy v_task from it and fine-tune on task demos, then add γ · (v_task − v_ref) to the base’s velocity field at every denoising step. Because the proxies share initialization + architecture, the residual isolates the task-induced distribution shift the base would have undergone if it had been fine-tuned — enabling 0-to-1 gains on tasks the base never solves (unlike RL²-VLA whose companion is bounded by base rollouts) while leaving the base weights and priors untouched (unlike LoRA which loses robustness under matched perturbations). Reports +53% absolute over π0.5 across 8 real-world tasks and +51% across 4 simulation tasks, beating LoRA, from-scratch specialists, residual policies, and DSRL on every simulation task. Bonus modality-routing extension: because v_task has its own encoder, extra modalities (point cloud, audio) can be routed into the RGB-only frozen base through the task proxy without retraining the base — a cheap sensor-addition story vs the unified-backbone-retraining answer of Embodied-R1.5 / Isaac 0.5.
• [2026-09-11] LIT — two-stage VLA/WAM training constrains visual conditioning via latent interface to prevent visual shortcuts: LIT (@djiafei thread) adds a training-time restructuring row distinct from Anchor-Align’s loss-side answer to the same shortcut-during-BC problem: Stage 1 trains the action expert on language + robot state + spatial goal without images; Stage 2 introduces vision through a supervised latent interface trained to recover the spatial goal — so any visual information reaching the action expert must first pass through a goal-relevant bottleneck. Cross-architecture LIBERO-Plus average lifts (π0.5 68.97 → 79.67%, MolmoAct2 63.62 → 71.92%, FAST-WAM 51.44 → 60.63%, ImageWAM 83.02 → 86.89%) with in-distribution LIBERO improving simultaneously across all four backbones. Real-world MolmoAct2 (300 demos, three tasks) lifts new-lighting 53.3 → 70.0%, top-camera-only 30.0 → 46.7%, distractors 50.0 → 63.3%, and in-distribution 74.7 → 88.0%. Trajectory attribution: distractors + blur leave LIT’s predicted trajectories near-original while baselines shift substantially; goal changes still produce expected trajectory changes — the intended causal signature of vision entering only through a goal-relevant bottleneck.
• [2026-09-10] Rhoda AI — Web-video pretraining scaling laws for robot policy learning (DVA): Rhoda AI publishes the first quantitative scaling curves for their Direct Video-Action (DVA) recipe on real hardware: web-video pre-training (no actions) + ~11 h teleop post-training on one industrial bearing-unpacking task raises at-speed completion 4% → 85% across XS/S/M/L, still rising at L. Pre-training compute helps most when robot demos are scarce (31-point gap at 12.5% demo data). Reinforces the DVA thesis that web-video pre-training transfers to real industrial manipulation without teleop-heavy scaling.
• [2026-09-10] SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators: SyncWorld demonstrates a training-free policy-improvement loop stacked on top of an existing VLA: the VLA proposes K action chunks, SyncWorld imagines each rollout from multiple views, and a VLM ranks them against the language instruction — the top candidate is executed, with neither the policy nor the WFM updated.
• [2026-09-10] Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness: Wenli Xiao demonstrates GPT-6 Astra doing first-pass physical ICL — a human-video demonstration dropped into the codex app is enough to drive a real tabletop arm via a motion-planning tool harness (Astra emits EE targets, harness runs IK). Pushes the LLM-as-controller row one step further than thijs — Astra paints the Golden Gate Bridge with a real robot arm and Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK: the demonstration channel is a pre-recorded human-body video (not a text prompt or in-scene physical demo) and the retargeting from human motion to robot motion happens entirely inside the LLM’s forward pass on top of a thin IK tool — no VLA, no fine-tuning, no policy training. Qualitative single-attempt claim only; contact-establishment regime not tested where prior Robocurve runs and Malik’s sensorimotor counter-position predict LLM+IK to struggle.
• [2026-09-09] Robot-Use Agents: Phillip Isola names the paradigm — “robot-use agents”, cloud LLM as puppeteer of a robot body — and stakes a positional counter-claim to the on-device customized-VLA line: per-robot performance stays below dedicated policies but rate of intelligence diffusion across the whole ecosystem is what changes, since the model runs in the cloud with no per-embodiment engineering. Names latency + safety-critical reliability as the surviving limitations. Reframes the ongoing empirical Astra/Fable/Claude-as-controller thread (Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK, Comparing Fable 5.1 to Fable 5 on robotic manipulation, thijs — Astra paints the Golden Gate Bridge with a real robot arm) as evidence of a deployment-trajectory shift, not a capability-ceiling one — and sits in productive tension with the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position (measures the ceiling) and Going Beyond World Models & VLAs Florence’s “video pretraining is a crutch” framing.
• [2026-09-08] EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data: EgoWAM (Georgia Tech RL2; Danfei Xu) fits the “two-headed VLA where world prediction shapes the trunk” recipe — HPT backbone + shared learnable action and future-token banks + flow-matching action head + swappable world-model head, trained jointly with L_BC + 0.05 · L_WM and world-model head dropped at inference so deployed latency matches a same-size BC policy at 30 Hz. Position: under BC co-training human/robot (and even human/human) embeddings sit in isolated clusters when actions do not align, while WAM co-training pulls them into a shared space via dynamics supervision, letting large-scale in-the-wild egocentric human data shape the shared backbone even when its action labels don’t transfer.
• [2026-09-08] thijs — Astra paints the Golden Gate Bridge with a real robot arm: Another five-day-after-launch LLM-as-controller data point on the recipe-lever board — GPT-6 Astra drives a real tabletop robot arm + paint brush + camera end-to-end (no VLA, no fine-tuned policy, no teleop) via a plan-one-minute + monitor-and-adjust control loop with natural-language human critique between attempts. Qualitative-only companion to the measured Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK Astra-vs-Fable-5.1 YAM-arm evaluation and to Sentdex — general multimodal LLMs may sidestep VLA/WAM research for high-level robotics intelligence‘s open-weights XGO-quadruped demo; artistic framing (“algorithmic art”) sidesteps the contact-rich functional-manipulation regime where prior Robocurve runs show LLM controllers struggle most.
• [2026-09-08] EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents: EmbodiedSkills (Wang, Zhang et al., Zhejiang + collaborators) adds an interface-typed row to the wrap-frozen-VLA-with-orchestrator cluster: rather than the prompt-engineered guides of Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA or the ad-hoc verify-and-recover step of Addressing the Orchestration Gap in Generalist Robots via Physical Agency (Pigey) Pigey, every skill decision goes through a shared executable-skill interface with prerequisite check + bounded low-level VLA execution + post-action verification hooks — so the low-level VLA (OpenPI/π0.5) can be swapped or task-adapted without touching the agent loop. Reports 86.20% avg on 50 RoboTwin 2.0 tasks and 97.40% across the four LIBERO suites with Qwen3-VL as the high-level model, plus a useful counter-signal at 12.5% on four memory-dependent RMBench tasks — evidence orchestration alone does not solve the memory-dependent long-horizon regime.
• [2026-09-08] OpenWAM — Extensible framework for video-action world models in robot learning: OpenWAM (Stanford SVL, AGPL-3.0 v0.1.0 pre-release) ships a training + evaluation framework for WAMs / video-action models with three maintained topologies (parallel_stream, dual_expert, causal_video_prediction), six standard video/action program selectors plus GJD, and a role-scoped extension SDK — sibling in the training-to-deployment authoring axis to FluxVLA Engine — Standardized VLA Engineering Foundation FluxVLA Engine but scoped narrowly to WAMs and to controlled program ablations across a fixed visual stack. The parallel_stream runtime is explicitly LingBot-compatible (Causal World Modeling for Robot Control (LingBot-VA)) — reads as an open academic-lab substrate for the WAM design space the closed industrial WAM line (LingBot-VA / Dyna-2 / µ₀) has been staking out. No benchmark numbers or checkpoints at filing time; framework only.
• [2026-09-06] How we actually test a robot hand: RLWRLD (Korea, WEF Technology Pioneer 2026) frames RealDex1 as an explicitly hand-agnostic VLA foundation model — hardware-agnostic intelligence layer fusing vision + tactile + force + torque + motion across a wide variety of dexterous hands — and pairs it with the All Hands Up! archive + private DexBench simulator to characterize the hardware-side failure surface across 10+ commercial hands. Reinforces the Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) full-stack thesis but stakes the opposite end-effector position: don’t ship a single reference hand, characterize the vendor market and train one policy that spans it.
• [2026-09-06] Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design: Transformer Transformer (Ha, Liu, Song; CoRL 2026) inverts the recipe-lever board: keep the task fixed (UMI end-effector trajectories + user reward) and optimize the embodiment as the decision variable, with the same DiT acting as embodiment generator, dynamics critic, and cross-embodiment controller under three masking schemes. RoboTokens tokenize any articulated rigid robot as typed continuous tokens (links, joints, motors, states, actions) with ID-based cross-references; jointly diffusing embodiment + state + action gives GPU parallelism across batch and timesteps so 128 candidates diffuse+evaluate faster than CMA-ES evaluates 5, and beats CMA-ES on speed and task performance across fixed-base + mobile-bimanual (Fig. 5). Dynamics Self-Guidance backprops reward gradients through the noisy DiT prediction into embodiment tokens (classifier-guided DDIM) — most valuable at low sample budgets (+242 to +341 reward on multi-trajectory bimanual vs Zeroth Order at N=1). Cross-embodiment control via the same DiT self-validates diffused designs at Pearson r=0.53 vs the 128-RL-expert baseline — early evidence a unified controller can substitute for a per-variant RL expert bank. Real hardware: fabricated ALOHA optimized for cloth-unfolding (Flingbot) hits 30% max-joint-speed reduction (8.9→6.2 rad/s) and 73% tracking-error reduction (22→6 cm) vs original ALOHA. Complements policy-side recipes (π*0.6, XR-1, Embodied-R1.5) on an axis they don’t touch — makes the embodiment an optimization target rather than an assumption.
• [2026-09-06] TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation: TAM (Son, Shkurti, Lee, Shah, Kim, Fox; CoRL 2026) stakes out the sharpest below-the-policy sim-to-real adaptation recipe on the board: a residual-torque module trained 100% in randomized simulation with analytic inverse-dynamics labels, sitting between the low-level controller and the robot’s torque interface. Same TAM weights drive an RL vision-based pushing policy (76.2% vs SysID 57.1%, RMA 52.4%), a BC flip policy (72% vs SysID 34%), and an MPC ball-on-plate controller (6.20 vs 4.80 goals/run) on a real Franka Panda with no policy retraining. Direct counter-position to RMA (which requires retraining every new policy with the adaptation branch) — moves the adaptation lever to the torque level so it’s plug-and-play under any policy, even closed-source ones. Sibling in framing to ReForce — force-aware retargeting via closed-loop residual force tracker ReForce (residual force tracker below ACT) but with the stronger policy-agnosticism claim: no policy observations, no rewards, no action-space assumptions.
• [2026-09-05] NAC: Neural Action Codec for Vision-Language-Action Models: NAC (Jawaid, Xiang; UTD) adds a learned-action-tokenizer row orthogonal to the recipe/backbone/data axes: port the neural audio codec stack (SEANet + multi-scale RVQ + Vocos/ISTFT decoder + DAC adversarial discriminator) to action chunks flattened as 1D multichannel waveforms, dropping the mel-spectrogram loss (which drives downstream success to ~0%) and keeping the discriminator (removing it also drives success to ~0%). Reports first place across LIBERO-10 (49.7%), RoboMimic (33.9%), and 8-task real-world (50%) vs FAST, OAT, VQ-VLA, Diffusion Policy, and binning at 12 tokens per chunk (matches OAT/VQ-VLA count, 3× fewer than FAST) and 3.5 ms end-to-end. Sibling in structural-interface framing to Action-QFormer Action QFormer: Structured Representation Shaping under Action Supervision in Vision-Language-Action Models but at the tokenization layer rather than the query interface; direct counter-position on velocity to Is Diversity All You Need for Scalable Robotic Manipulation? (Shi et al. normalize velocity away in data; NAC preserves it in the tokenizer’s high-frequency residuals).
• [2026-09-05] RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning: RoboTok (Rice RobotPI Lab; Qian et al.) supplies a data-side lever for VLA pretraining: a motion-similarity retrieval index over web human-manipulation video, keyed on 3D hand trajectories in an actor-centered reference frame. Instead of a fixed curated corpus, retrieves demonstration videos matched to a single query motion — the concrete “what to align on” answer to MimicLabs’ What Matters in Learning from Large-Scale Datasets for Robot Manipulation finding that targeted alignment beats scale in the OXE-scale regime.
• [2026-09-05] Jay Chooi (Robocurve) — Astra vs Fable 5.1 on real-robot manipulation via IK: Robocurve pits GPT-6 Astra against Fable 5.1 on YAM-arm manipulation via IK — Astra matches success rate with 3.9× fewer tokens and 1.6× lower cost, extending the pattern of general LLMs as competitive robot controllers.
• [2026-09-04] Comparing Fable 5.1 to Fable 5 on robotic manipulation: Fourth Robocurve real-hardware Inspect Robots eval — first at 20-rollouts-per-condition (80 total) and first on the end-effector-pose + IK-solver interface rather than raw joint targets. Claude Fable 5.1 vs Fable 5 on YAM-bimanual block-into-bowl (40% vs 5%, 1.49× fewer tokens) and puzzle-into-groove (10% vs 0%, 1.55× fewer tokens); Fable 5 stalls at first contact 13/20 times on the bowl task (contact-establishment failure) while exhausting its 20-LLM-call budget in 18/20 trials, so the gap is partly budget-confounded but the failure structure matches the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position’s specific prediction for LLM-as-controller with IK.
• [2026-09-04] ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine: ACE-Data-0 (ACE Robotics + S-Lab NTU) releases the multimodal pretraining substrate for ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining‘s VLA recipe — 150+ hr of goal-directed household activity with synchronized ego + 8-view exo video, whole-body + hand mocap, object 6-DoF, tactile, and audio on one clock, gated to non-commercial academic access.
• [2026-09-04] Sentdex — general multimodal LLMs may sidestep VLA/WAM research for high-level robotics intelligence: Sentdex posts a self-contained thesis + working zero-shot demo: GLM 5.3 Flash (and reportedly DSV4F + Qwen 3.8 27B) controls an XGO mini quadruped+arm via its own vendor SDK as tool calls over camera frames, no VLA / no fine-tuning / no teleop data — sibling framing to Lerrel Pinto — in-context learning for robots is not that hard and LeRobot demos Claude Code operating SO-ARM101 zero-shot via Anthropic MHS — self-calibration to 4.1mm accuracy but with open-weights frontier LLMs and no eval harness in the loop, adding a fresh cell to the LLM-as-controller row of the recipe-lever board. Explicitly does not claim locomotion is covered.
• [2026-09-03] Getting a Grip on Robotic Data Collection — Co-Designed Koala Gripper for Force-Mediated Manipulation: RAI’s post frames the reactive-VLA problem crisply: because today’s policies observe the current moment without memory, the instantaneous observation must fully determine task state — force and proprioception directly resolve ambiguities vision alone cannot, and this is why co-designed handheld collection matters more than for higher-scale-lower-fidelity substrates. Argues force-mediated manipulation (assembly, disassembly, tool use) is the specific task class where current VLAs are structurally blocked, not just data-limited.
• [2026-09-02] C³ache: Accelerating World Action Models with Cross Inference Chunk Cache: Motivates cross-chunk residual caching by contrasting VLAs and WAMs: VLAs generalize only to the behaviors demonstrated in action-labeled data while WAMs learn from unlabeled video and generalize further at higher inference cost — this paper closes ~2× of that inference-cost gap training-free by caching DiT residuals across action chunks.
• [2026-09-02] TwinDEX: A Scalable Path from Robot-Free Data Collection to Real-World Dexterous Manipulation: TwinDEX (X Square Robot) reports a wearable-only-data, zero-on-robot-data cell on the recipe-lever board at dexterous scale: bimanual three-finger nine-DoF wearable twin of the robot end-effector, few-hundred demos, 20+ step chemistry experiment with tool switches, fine force control, and bimanual coordination completed after training from scratch, plus 5.3× effective data throughput vs teleop in multi-task eval. Announcement surface only — no backbone, action head, per-task success rates, or head-to-head with π0.5 / GR00T / other filed VLAs disclosed. Sharpens the “make the collection tool embodiment-invariant” position (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic, HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet) HandUMI, Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1) by dropping the post-training on-robot leg entirely rather than reducing it.
• [2026-09-01] What Matters in Learning from Large-Scale Datasets for Robot Manipulation: MimicLabs (Saxena, Bronars, Mandlekar, Xu) provides the first filed dimension-level breakdown of what dataset composition axes drive VLA co-training gains on DROID. Camera-pose alignment + broad spatial-arrangement diversity in the retrieved subset beats naive DROID co-training by up to +70% on real-robot target tasks; naive full-corpus co-training can underperform target-only training when camera poses are misaligned. Direct actionable retrieval recipe for anyone building a target-task VLA on top of DROID/OXE.
• [2026-09-01] The Imitator Game: Benchmarking Robot Imitative Ability Beyond Action Prediction: The Imitator Game (Zhou et al., HKU + collaborators) supplies the diagnostic axis this concept has been missing: a four-level (L0–L3) benchmark that progressively widens the gap between the human demonstration and the robot’s own scene, plus IG-10K (20,000+ paired human-robot episodes across 50+ tasks, real+sim, all four levels) and the Imitator Arena for blind A/B human evaluation. Load-bearing empirical finding across 9 SOTA models / 15 trained variants: performance is stable L0→L2 then collapses at L3, identifying functional substitution — same intent through a different object affordance — as the decisive barrier that current recipe levers (action-pretraining, clean teleop, unified-VLM pointing, WFM+action-expert, ICL) fail on. Human-video-conditioned models beat caption-conditioned ones under matched protocol; every model falls <13% zero-shot on 5 unseen tasks; IG-10K-pretrained fine-tuning on only 10 paired human-robot demos yields large gains that grow with pretraining scale. First filed benchmark to level-grade the human-to-robot imitation gap rather than aggregate over it.
• [2026-09-01] AttenA+: Rectifying Action Inequality in Robotic Foundation Models: AttenA+ (Peng, Ma, Cao et al.; HKUST-GZ + HKU + IDEA) adds a loss-side reweighting row to the recipe-lever board that is training-side and orthogonal to backbone / data / RL choices: compute per-timestep end-effector velocity from the ground-truth action, invert it into a per-token attention weight (clipped to [1, C] with C ∈ {2, 3} load-bearing for stability), and use it to scale either the discriminative regression loss or the flow-matching objective — so low-velocity precision-critical segments (grasp, align, place) get more optimization capacity than high-velocity transitional segments. Reports OpenVLA-OFT 97.10 → 98.60% on LIBERO overall (+1.5 pp, +2.1 pp on LIBERO-10 long-horizon), π₀ 96.85 → 97.95% (+1.10 pp) — paradigm-agnostic to discriminative vs flow-matching heads — Fast-WAM 91.80 → 92.46% on RoboTwin 2.0 without embodied pretraining (edging out embodied-pretrained LingBot-VA 92.2%), and real-world Franka 92.5 → 97.0% across 4 tasks × 50 trials with the largest gains on multi-object and long-horizon. Sibling and direct counter-recipe to Is Diversity All You Need for Scalable Robotic Manipulation? on the same velocity-as-signal axis: Shi et al. canonicalize velocity by resampling demos to a canonical chunk length (velocity is noise to normalize away), AttenA+ preserves velocity and upweights low-velocity tokens in the loss (velocity is structure to keep). Ablations: no single velocity → weight non-linearity dominates all task types; C = 5 degrades; explicit limitation is that velocity-as-criticality inverts in ballistic tasks (high-speed grasping, throwing).
• [2026-09-01] LightNav-0: Scaling Real2Sim2Real for Zero-Shot Generalist Navigation: LightNav-0 (Light Origins) is the second filed VLN system (after Robostral Navigate: single-camera AI navigation Robostral Navigate) to put quantitative numbers behind unified single-camera language-first navigation, and sharpens the Robostral recipe on three axes: a two-point Point-CoT reasoning trace (object + affordance image-space points) instead of a single pointed waypoint, a three-level RVQ tokenization of ten-waypoint SE(2) trajectories (0.72 cm reconstruction, 3× 256-entry codebooks) instead of continuous waypoint + fallback metric displacement, and Stage-III GRPO on discrete action tokens instead of CISPO on a continuous head — sidestepping the flow-matching-VLA workarounds (RECAP CFGRL, VINE) the manipulation side has been forced to invent. Ranks first on all ten monocular head-to-heads across VLN-CE / Matterport3D / HM3D / HM3D-OVON / EVT-Bench and first in every INSIGHT-Bench instruction × scene cell (7 open-source policies, matched forward-RGB protocol); same checkpoint transfers zero-shot to humanoid / quadruped / aerial / wheeled robots and to non-robot domains (Minecraft, first-person games, in-the-wild egocentric video). Adds a “navigation-VLA” row to the recipe-lever board with an ER-before-A staging (LightNav-ER: 67.4 macro-avg on 8 embodied-reasoning benchmarks before action alignment) that mirrors Gemini Robotics-ER / MolmoER on the manipulation side. Model + code + tech report open.
• [2026-08-31] Zetta ζ: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence: Zetta (Tsinghua AIR + MSR Asia; Ding et al.) stakes out the highest-frequency, no-LLM-in-inner-loop cell of the wrap-frozen-VLA-with-companion pattern — code-based runtime critics and recovery skills execute above the base policy’s action rate, so intervention latency does not inherit VLM or agentic-LLM cost. Three time-scale-separated loops factor governance (action-rate critics) from proposal (rollout-level critic-recovery candidates) from promotion (validation-gated skill updates via a paired same-seed gate + held-out seeds 1..20). Reports LIBERO-Pro 90.8% and RoboCasa 93.6% under the paper’s rollout budget with 11.1× inference speedup vs comparable agentic baselines; success continues to scale with self-exploration experience and learned critic-recovery skills transfer zero-shot across rollouts. Z-Infra decouples the harness from execution resources so the same code runs across LIBERO, RoboCasa, and per the roadmap NVIDIA Cosmos / RoboTwin / ManiSkill / BEHAVIOR without changes. Distinct from FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models FailSafe (10-step polling companion VLM emitting deltas), Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA (LLM planner at subtask cadence), and Enact launch — post-training infrastructure that generates targeted recovery data for robotics VLAs Enact (offline DAgger + human-in-the-loop) in intervention frequency and validation discipline.
• [2026-08-31] Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models: INDI (Lee, Mo, Han) adds an intent-distillation auxiliary-loss row to the recipe-lever board distinct from the future-prediction / value-conditioning / world-model / representation-anchoring rows: a frozen teacher VLM interprets the full demonstration segment (observation + instruction + coarse action summary + execution video) into a multimodal behavior-level intent latent, and the action decoder is trained to recover that latent at an intermediate layer from present-only inputs — no teacher at deployment. Reports GR00T-N1.7 SimplerEnv-Bridge 64.3 → 84.7% (+20.4 pp), RoboCasa Kitchen 64.1 → 70.3% (+6.2 pp), and real-world 62.0 → 68.7% avg with up to +12 pp on longer-horizon tasks; consistent gains on π₀.₅. Distinct from Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment Anchor-Align’s per-frame VLM representation anchoring by targeting a behavior-level temporal-semantic abstraction, and from LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID by keeping intent as an internal representation-shaping signal inside a single VLA rather than an external interface between pipeline halves.
• [2026-08-31] Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models: VLAct (Yang, Wang et al.; CUHK MMLab / SmartMore / HKU) stakes out a representation-centric continued pre-training row distinct from Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment Anchor-Align’s fine-tune-time loss-side answer and OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence DM0.5’s separate-LR-group answer: freeze the vision encoder and lower half of the LLM during continued pre-training with caption supervision anchoring the frozen half; jointly supervise the trainable half through OFT + PI + GR00T action heads that are discarded at the end of pre-training (forcing head-agnostic representation); unify cross-embodiment action spaces via shared coordinates + masked inactive dims + wrap-aware periodic-joint loss. Reports 82.6% LIBERO-Plus Total (+7.6 pp over matched Qwen3VL-OFT baseline), 92.5%/90.8% RoboTwin Clean/Randomized under Data Scaling, and exceeds 100%-data baselines with only 20% of RoboCasa-GR1 fine-tuning trajectories on an embodiment held out from continued pre-training. Continued pre-training runs on 16 GPUs against open datasets (DROID + MolmoAct + InternData-A1 + RoboCoin), backbone StarVLA/VLAct_Qwen3_Pretrain and four downstream checkpoints released under MIT. Preprint (Aug 2026), no arXiv id at filing time.
• [2026-08-30] TemporalFlow-VLA: Learning Physically Grounded Execution History for Long-Horizon Robot Manipulation: TemporalFlow-VLA (Yang, Lu, Su et al.) stakes out a training-only physically-grounded supervision row on the VLA-history axis distinct from StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models StreamPI (attention-mask + interval schedule, zero new params, no new supervision) and from MEM: Multi-Scale Embodied Memory for Vision Language Action Models MEM / BridgeVLA++: A Data-Efficient, Generalizable, and Memory-Augmented Vision-Language-Action Framework for 3D Manipulation BridgeVLA++ (dedicated memory architectures): construct a robot-surface temporal flow target offline from robot states + geometry + calibrated cameras, supervise two execution-aligned temporal queries against it, and drop the geometric pipeline at deployment so the queries feed structured history to the action expert with no motion estimation at inference. Reports 97.63 ± 0.26% average LIBERO (96.60 ± 0.87% LIBERO-Long) and 85.5%/84.2% Clean/Randomized on 12 RoboTwin tasks with the clearest margin on longer-horizon multi-stage manipulation; controlled history-intervention analysis shows action prediction depends on both historical content and temporal order. Asynchronous feature caching keeps server-side sampling latency at single-frame level. Sharer framing (“useful for sidecars”) points at the deployment-time property that the geometric supervision path is training-only.
• [2026-08-29] Inspect Robots — An Open-Source Evaluation Framework for Physical AI (Robocurve): Inspect Robots (Robocurve, MIT) is now the third-party audit substrate that ships both --policy agent (LLM-as-controller through motion tool calls) and --policy xpolicylab (40+ served VLAs including π0/π0.5, GR00T, OpenVLA-OFT, RDT-1B, SmolVLA, ACT) as swappable policies against the same embodiment, with a check_compatibility step that fails fast before actuation on mismatched action spaces or control rates. Turns “swap in a new frontier LLM vs a trained VLA on the same YAM arm” from a config-editing project into a single flag change, and formalizes the small-N-per-model rollout methodology behind the three prior Chooi/Robocurve real-hardware tweets into a schema-versioned EvalLog per run.
• [2026-08-29] Lerrel Pinto — in-context learning for robots is not that hard: Lerrel Pinto (NYU) short-form endorsement that ICL for robot policies is achievable without exotic recipes, joining the S1 / GEN-1.5 / HOST chorus — a claim without evidence, but from a lab that’s an active player in the space.
• [2026-08-28] OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence: DM0.5 (Dexmal; OpenDM release) stakes out a “keep VL capability strong while learning to act” recipe row distinct from π*0.6’s RL-on-flow-matching: Gemma3-4B VLM + 680M flow-matching Action Expert with separate learning-rate groups (VLM at smaller LR) plus an internal automatic VL data generation pipeline (image + video + visual instruction, targeting scene semantics / physical dynamics / task progress) to fight catastrophic forgetting during action co-training — the LR-group + synthetic-VL knob-set alternative to Anchor-Align’s distillation loss. Long visual history (~60 s per external coverage) enables in-episode learning from a start-of-episode human demonstration and mid-episode camera-perturbation adaptation without a reset. Reports 99.0 LIBERO, 93.6/93.3 RoboTwin2.0 Clean/Rand, 89.0/53.6/44.1 VLA-Arena L0/L1/L2, 19.34% RoboDojo-Sim SR (ARX X5 bimanual via MEM variant), and 54.42/43.0% Score/SR on RoboChallenge Table30V2 real-world (vs π0.5 31.48/14.3%). Widest margins over π0.5 at real-world Table30V2 and on the harder VLA-Arena tiers, matching the “long context + LR-group” bet’s predictions; Twitter framing is “pi0.5 of China”.
• [2026-08-28] LeRobot demos Claude Code operating SO-ARM101 zero-shot via Anthropic MHS — self-calibration to 4.1mm accuracy: LeRobot demos Claude Code operating a real SO-ARM101 with no policy, no teleop, no demonstrations — agent self-calibrates via 16-dot torque-off gripper probing, reports 4.1mm position accuracy — a live counter-example to the “manipulation requires trained policy” framing.
• [2026-08-27] Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning: Q-Planning (Giridhar, Khandelwal, Collins, Georgiev, Garg) adds a Q-function-only self-improvement row to the recipe-lever board that keeps the multi-billion-parameter BC policy fully frozen and puts all the deployment-time learning into a small off-policy Q-function used to re-weight N BC candidates at inference. Distinct from π*0.6: a VLA That Learns From Experience (RECAP) RECAP (which folds the value signal into the BC policy via advantage conditioning + CFGRL) and from VINE: Taming Generative Control Policies for Reinforcement Learning VINE (end-to-end value-gradient BPTT through flow-matching denoising steps) by refusing to touch the policy weights at all — the load-bearing structural asymmetry per the paper is that failed rollouts are legitimate training signal for a Q-function (low-value states) but actively harmful supervision for BC. LIBERO-10 93→99%, RoboTwin 83.8→91.4% over 10 iterations; on two real-world contact-rich bimanual tasks with BC frozen, five iterations lift stack-cups 40→90% and insert-wallet 25→80%, whereas SFT-on-successes stalls at 55% and 30%. Reported as the only method among Best-of-N / filtered SFT / IBRL / DSRL / DAWR that improves stably from failures without training an auxiliary actor at matched online budget. Author tweet-thread also documents the 25% BC ceiling on 100 teleop demos of the harder task as the “BC by construction” problem this paradigm attacks.
• [2026-08-27] Previewing the Model Hardware Standard: Anthropic’s Model Hardware Standard (MHS) research preview stakes out an architectural alternative to the VLA layer: instead of a neural policy mapping perception+language to actions, the agent controls the device driver directly through read/write primitives + natural-language device tags, with vendor buy-in from Universal Robots, Doosan, Hugging Face LeRobot, and Raspberry Pi. Complements the Claude-as-VLA-supervisor thread (How Claude Performs on Robotics Tasks) by treating the device itself as the tool one abstraction layer below the VLA — relevant for hardware that has no VLA policy behind it.
• [2026-08-27] V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning: V-JEPA 2-AC — freeze a large action-free video backbone, train a 300M block-causal predictor on 62 hours of DROID, run image-goal MPC in latent space on unseen Franka arms; a cheap alternative to full diffusion-rollout VLA pipelines.
• [2026-08-27] Behavior Prompting Policy: Demonstrations as Prompts for Manipulation: Behavior Prompting Policy (Patel, Pekarek, Castro Hernandez, Song; REAL Lab Stanford; arxiv:2606.30457, project page behavior-prompting.github.io) adds a mechanistically-ablated in-context visuomotor architecture row to the recipe-lever board, sibling to S1 and GEN-1.5 on the ICL-native axis but distinct in staking the position on specific ingredients rather than pretraining scale: attention-pooled per-timestep prompt chunks (observation + proprioception + actions merged into one embedding) cross-attended against the current observation, with the diffusion action decoder decoupled from the prompt encoder so multiple denoising steps do not re-reference the full prompt. Reports 80.7% error reduction vs Goal-Image and 33.3% vs ICRT on 50 unseen DrawAnything drawings; matches finetuned π0.5-family baseline on LIBERO-Gen Combination/Chain without foundation pretraining; +20.8 pp over Language conditioning in an ablation regime where 2nd-step tasks are removed from training on LIBERO-Gen Chain (vs +10.7 pp with 2nd-step tasks). Diagnoses ICRT’s full-rollout-history context as OOD-brittle via spurious correlations — the sharpest filed critique of the transformer-context-with-history recipe. Explicit failure mode: low training task diversity weakens BPP task conditioning vs language (3-task laundry-folding case study shows wrong-fold and hesitation failures), and BPP has not been shown to enable one-shot execution of entirely new action primitives — only compositional new tasks over seen primitives. Ships iPhUMI: iPhone-based UMI extension with ARKit-based instant SLAM (no environment mapping step) and wireless prompt streaming from the same collection device.
• [2026-08-27] Project SuperDex — A unified simulation platform for dexterous manipulation: Project SuperDex (Meta) is a dexterous-manipulation-focused simulation + RL platform — infrastructure play for the sensorimotor-heavy end of the VLA design space, with tactile sensors and non-convex contact as first-class solver features and a Gymnasium-compatible RL layer in early preview.
• [2026-08-26] Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment: Anchor-Align (Dalal, Patel, Jain et al., UIUC + Meta FAIR; arXiv:2607.13429) adds an on-observation representation-anchoring row to the recipe-lever board: keep the VLA fully trainable and pin its VLM representation via a layer-wise distillation loss against a frozen copy of the same VLM computed on the robot observations themselves, paired with a language-action alignment loss that discretizes actions into motion-direction words and jointly trains language + action prediction on the same observation — closing the observation-mismatch gap that co-training on separate image-text data leaves open. Distinct from Harness VLA / LFP (freeze the VLA entirely) and Action QFormer (structural-interface change): Anchor-Align is the loss-side answer that leaves the interface unchanged. Real-robot xArm7 lifts both applied VLA backbones (28→54% and 37→60%), and beats every LIBERO-PRO / LIBERO-Plus baseline including Co-training + Knowledge Insulation — most decisively on LIBERO-PRO Position Swap (22.6 vs 2.3 for VLA-Adapter and 0.0 for the other four baselines), the memorized-trajectory-trap regime. Sharer notes anchoring is the load-bearing component per ablation and Anchor-Align is “almost always” better than KI.
• [2026-08-26] Isaac 0.5: An Open-Weight Embodied Foundation Model: Isaac 0.5 (Perceptron AI) stakes out a general-action-free-video-as-primary-scaling-axis row on the recipe-lever board: 36B sparse MoE with null-expert routing (0–8 of 256 experts per token) plus a proprietary self-supervised future-percept objective, trained on 1,000,000 h of general video + 375K h egocentric + 375K h UMI + 100K h robotics. Reports a within-grid teleoperation crossing moving from 5,884 h → 28 h (210.3× reduction) as general video scales 1,000 → 1,000,000 h at fixed 80:30:30 composition, bounded 83× to 300× by adjacent measured rungs. Empirical loss-vs-video slope stabilizes at −0.21 loss per 10× video from ~100 teleop-h onward and is nearly flat at 1 teleop-h — video and teleop are complements at the low end, substitutes at scale. Sharpest quantitatively-calibrated counter to Going Beyond World Models & VLAs Florence’s “video pretraining is a crutch” thesis and to GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5’s from-scratch-real-interaction bet, and the first filed instance where general action-free video (not egocentric or UMI) is the primary scaling axis. Open weights + inference stack + 35+ per-embodiment deployment configurations released. No headline benchmark comparison in the released body text.
• [2026-08-26] Behavioral Cloning Mystery: Seohong Park (ex-OGBench, now in robotics) reproduces four real-robot BC mysteries — overfitting is often beneficial and 50K episodes can lose to 10K on the same distribution, closed-loop $\pi(a_t\mid s_t)$ fails outright while open-loop 25-step chunked policies succeed with infinite data (history-conditioned closed-loop is worse, not better), state-based single-task pick-and-place needs $[4096]\times 8$ residual MLPs (~0.5B params) to work well, and hand-scaling object $xyz$ ×10 beats standardized features despite identical train-time flow-loss/MSE — inside a purpose-built MJWarp benchmark whose scripted demonstrators mimic human-demo properties (narrow support, non-Markovian, smoothly random via Hermite splines) and stream fresh data consumed exactly once. All four mysteries reproduce under state-based flow BC with no vision encoder, so they are properties of the data + BC objective, not perception; phenomenological complement to Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control? Lazzati et al. and Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies Zeng et al., strengthening both by ruling out perception as a confounder. Park’s thesis: all four are surface expressions of test-time distribution shift + expressivity mismatch between non-Markovian data and Markovian(-chunked) policies; LLMs escape this only because internet-scale pretraining makes essentially every test-time input in-distribution.
• [2026-08-25] Figure launches Index — crowdsourced app-based real-world video dataset for humanoid robots: Figure launches Index — app-based paid-crowdsourced data-collection pipeline for training Helix VLA at industrial scale (16M+ videos, 264K downloads, 108 countries, $15M paid out); Figure commits$1B/12mo to data + compute and targets 100× scale-up.
• [2026-08-25] Legato: Learning Native Continuation for Action Chunking Flow Policies: Legato (project page only; no algorithmic detail at filing) adds a fourth training-time entry on the async-inference / chunk-boundary axis alongside REMAC (Real-Time Robot Execution with Masked Action Chunking), πR² (πR²: Reactive Real-time Flow Policies), and A2A (A2A: Action-to-Action Flow Matching) — per the sharer’s framing, fine-tunes the flow-matching policy to natively continue an executed action prefix rather than treat it as an inpaint condition at sampling time (RTC’s approach). Reports lower completion time and roughly-halved chunk-overlap RMSE + NSPARC vs RTC on five real-world tasks (Bowls / Pour / PickPlace / Drawer / Towel) under matched pretraining + dataset + hyperparameters, and beats Training-Time RTC on Pour specifically (9.72 vs 9.46 score, 75.7 s vs 81.7 s completion, NSPARC 1.65 vs 2.46). Method opaque without accompanying paper.
• [2026-08-25] Sergey Levine: Current State of Humanoid Robotics, China & Future Predictions: Sergey Levine (Physical Intelligence co-founder) on The Peterman Pod frames the hardware supply chain — reliable low-cost actuators/sensors/mechanicals, largely sourced from China today — as the binding constraint on US robotics progress, an outside-the-model-recipe axis that sits next to Humanity's Last Machine: A Deep Dive on Humanoid Hardware‘s BOM analysis and complements the software-side levers this concept tracks. Transcript unavailable; summary reconstructed from external coverage of the same episode.
• [2026-08-21] Robo Robotics launch — Robo-T bimanual humanoid, sub-$10/hour with Roboport teleoperation platform (Kyle Noble): Robo Robotics launched publicly with Robo-T — a bimanual humanoid (2× 6-DoF arms + grippers, 10 lb, one-human footprint, sub-10/hour deployed) plus **Roboport**, a teleop platform that unifies demonstrations, live shift operation, and DAgger-style interventions into a single closed-loop training substrate. Third recent vertically-integrated launch (sibling to [[tweet-iliraliu-2026-08-11]] MicroFactory and [[tweet-jameswstevens-2026-08-11]] Enact) that treats the teleoperator→supervisor→rare-intervener transition and the per-intervention data flywheel as the core R&D axis rather than the model layer. Launch essay explicitly frames the current 10–30/hr teleop labor + 1:2 or 1:3 robot:operator ratios as existential and models the deployment-economics problem in Waymo-style rare-intervener terms. Announcement + positioning surface only — no per-task numbers, no intervention-rate data, no BOM.
• [2026-08-21] T-Rex: Tactile-Reactive Dexterous Manipulation: T-Rex (Berkeley + NVIDIA GEAR; arxiv version of Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction) is the first filed VLA where a modality other than vision drives architectural choices — a Mixture-of-Transformer-Experts with three experts (latent, action, tactile) bound by asynchronous cascaded flow matching split at τ=0.4, with the tactile expert firing 4× per visuomotor tick and reusing cached VL context so per-step cost is dominated by lightweight tactile denoising. Extends the EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data three-stage recipe (22,889 h egocentric pretraining → 100 h tactile mid-training → skill-specific post-training) with the tactile expert grafted on during mid-training, reaching 65% avg across 12 contact-rich tasks vs 35% EgoScale, 17% π0.5, 15% Tactile-VLA. Reports the concept-relevant negative result that naive tactile bolt-on regresses performance (π0.5 + tactile 6%) — tactile integration is architecturally load-bearing.
• [2026-08-21] Anirudha Majumdar — pure zero-shot task inference is the stronger test of robot foundation models: Anirudha Majumdar (Princeton) proposes an evaluation-methodology ablation for one-shot ICL claims in robot foundation models — strip the physical prompt, place objects on a table, and measure pure zero-shot task inference; predicts non-trivial success rates for GEN-1.5-class models, which would separate demonstration-conditioned lift from pretraining-data-coverage.
• [2026-08-20] The Multimodal Intelligence of Muse Spark 1.2: Muse Spark 1.2 (MSL) enters the VLA lineage with a two-tier system: a specialized Muse Spark variant is the high-level planner + orchestrator (scene understanding, object disambiguation, subtask decomposition) and a Muse Spark VLA policy executes each subtask on a bimanual platform (brush disambiguation, lunch packing, object finding). First MSL VLA on file, and an instance of the “VLM-as-orchestrator over VLA” recipe alongside Harness VLA / VoLo / Claude+MolmoAct — with the twist that both the orchestrator and the low-level policy are specialized variants of the same base model. Announcement post; no architecture, params, benchmark numbers, or task success rates disclosed.
• [2026-08-20] Going Beyond World Models & VLAs: Pete Florence’s April 2026 positioning post is the philosophical prequel to GEN-1.5 and the sharpest meta-counter-recipe on the board — argues the VLA label itself is a category error alongside the world-model label, that ~99% of GEN-1’s params being from-scratch is a deliberate rejection of both VLM-backbone and video-model-backbone starting points, and that VLM/video pretraining is a “crutch” for robotics data scarcity that Generalist has now removed at >500K hours. Direct counter-position to π*0.6, Spirit-v1.5, Embodied-R1.5, XR-1, µ₀, and 1XWM at the framing level rather than the recipe level. Companion piece to Towards Machines with a Thousand Hands and GEN-1.5: Embodied Foundation Models are One-Shot Learners.
• [2026-08-19] GEN-1.5: Embodied Foundation Models are One-Shot Learners: GEN-1.5 (Generalist AI) reports emergent one-shot in-context learning of physical skills from a single 3–12 s sensorimotor demonstration inserted into a 30 s context window — 59% avg across 10 tasks with zero gradient updates, 83% after 10 gradient steps on 5 min (~50 demos), and 66.5% held-out success after 1 step on 1 min with <0.15% weight change. No meta-learning, no ICL-specific architecture, no simulation data in pretraining, no context-packing infrastructure — the capability is claimed as emergent from Generalist’s real-interaction pretraining engine at 8+ months of continuous training. Physical prompts compose (two demos → chained behavior with intermediate bridging motions), transfer sim→real (sim-recorded prompt drives real execution despite zero sim in pretraining), and cross the embodiment gap (human-hand demo through the robot’s cameras drives robot-hand execution). Adds an “engine-scale emergence” row to the recipe-lever board distinct from RECAP’s RL, Spirit-v1.5’s clean teleop, or Embodied-R1.5’s pointing/grounding lever — the claim is that no explicit lever is needed once pretraining scale + data quality clears a threshold. No shared-benchmark head-to-heads; tasks described as simple and short-horizon.
• [2026-08-19] Grip — waste-sorting robots picking deformable trash with foundation models (YC S26): Grip (YC S26) is an application-domain data point: waste-sorting startup betting that current foundation-model manipulation policies can already handle deformable, entangled, open-vocabulary trash on commodity hardware, with per-cell picks aggregated into fleet-wide retraining.
• [2026-08-18] Gemini 3.7 Flash saturates Robocurve physical tool-use benchmark at 92% vs 32% for 3.6 Flash (three weeks apart): LLM-as-controller ceiling is a moving target on 3-week timescales: Robocurve reports Gemini 3.7 Flash at 92/100 on a “cover a block with a towel” task with 9.0 s per-turn latency, vs Gemini 3.6 Flash at 32/100 and 17.5 s three weeks earlier — both controlling paired YAM arms via raw joint targets, no IK, no motion planner. Contrast with Gemini Robotics-ER 2 vs Claude Opus 5 clapboard eval — ER2 threw robot arms off table, hallucinated success (Robocurve / Jay Chooi)‘s 0/5 for Gemini Robotics-ER 2 on clapboard: general-purpose Flash refreshes are now outpacing the robotics-specific line on this harness.
• [2026-08-17] ReForce — force-aware retargeting via closed-loop residual force tracker: ReForce (@wu876768 tweet) proposes a residual-composition pattern for turning a learned VLA policy (specifically ACT) into a force-safe one: leave the VLA producing motion+force targets, add a sim-trained residual force tracker on top that consumes live tactile feedback. Reports 30%→70% force-safe success and 7/10→0/10 severe missing-contact failures on paper-cup grasping over the ACT baseline. Sits alongside FACT: Force-Aware Contact-Rich Manipulation via Timestep Modulation — demystifying when and why VLAs fail on contact-rich tasks FACT (per-layer force modulation of the VLA itself) on the axis of where force integration should live — inside the policy vs a downstream residual.
• [2026-08-15] FACT: Failure-Aware Causal Training for World-Action Models: FACT (Peng, Liang, Yan, Hansen, Wang; UCSD) stakes out the WAM-side answer to the failure-data-reuse question, distinct from RECAP’s policy-side advantage conditioning (π*0.6: a VLA That Learns From Experience (RECAP)) and FailSafe’s runtime companion-VLM (FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models): flip the WAM to action-first + video-second so failed rollouts become valid supervision at the world-model layer, and reuse the jointly-trained progress head as an inference-time action scorer. RoboTwin 50-task avg lifts 81.8 → 85.6 → 87.5% (no video co-training → FACT → FACT + failures); failure-fraction sweep 0/50/100% → 32.7/44.7/57.3% success (monotone scaling); real YAM bimanual seen 82→92 and unseen 67→82 with +failures + scoring — approaches Motus / LingBot-VA without matching data scale. Future-frame PSNR on failure futures 19.51 → 25.92 dB directly documents the “success-biased hallucination” mode that success-only WAM training induces.
• [2026-08-15] Simulation Pre-training for Dexterity (SPD): SPD (Isola / Finn / Karen Liu labs) adds a sim-teleop pre-training + history-conditioning row to the recipe-lever board: DiT policy with sliding-window causal attention (32-timestep window, matched rolling KV cache) trained on 75 h of MuJoCo VR-teleop with Muon at fixed 1e-3, then 1–2 h real teleop fine-tune per task. Sharpest empirical counter yet to π0’s w=1, c=32 design choice (π*0.6: a VLA That Learns From Experience (RECAP)) for dexterous contact-rich tasks: w=32, c=8 (32-step history + 8-step chunks) is the strongest configuration in both training regimes and gains +18 pp average task progress from pre-training vs ≤3 pp for other variants. Under w=1, shrinking chunks to c=8 collapses the policy (shaky, temporally incoherent) — history conditioning is load-bearing for short reactive chunks. Beats matched-arch BC on all five real-world bimanual dexterous tasks.
• [2026-08-14] Thea: Towards the Harness of Embodied Agents: Thea takes the same positional stance as Waddle — VLAs (and other robot capabilities) are tools the top-level LLM agent orchestrates, not the top-level policy. Distinguishes the pattern from end-to-end VLA training by isolating the two adapter primitives (Scene Graph as Context, Evaluation as Exit Codes) needed to make the coding-agent scaffold work on top of a VLA-as-tool substrate.
• [2026-08-14] H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models: H2R-Bench evaluates whether video WMs can bridge the human-to-robot embodiment gap and convert human manipulation observations into robot-centric training resources for downstream VLAs — an upstream diagnostic for the human-video-as-VLA-substrate recipes trending across this page.
• [2026-08-14] DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation: DreamX-Phi 1.0 takes actions (per-arm end-effector poses + gripper states) as geometrically-structured attention conditioning — per-arm SE(3) transformations injected via PRoPE-style relative encoding — rather than as scalar tokens or generic conditioning vectors, arguing this is what prevents the “wrong arm moves” failure mode when a video WM is used as a manipulation prior.
• [2026-08-12] Flex-π: A Multi-Stream World-Action Model with Compute Flexibility: Flex-π (UW + Allen AI; Yan et al.) adds a runtime-compute-flexibility axis to the recipe-lever board: one 6B checkpoint denoises RGB + 3D pointmap + DINOv3 semantic latents jointly with actions in a shared space, and per-stream latent dropout + cross-modality forcing during training yield 56 deployable input/output combinations from the same weights (fast action-only ~60 ms vs full joint ~193 ms on RTX 5090, a live speed–accuracy dial). Load-bearing training-recipe finding: cross-modality forcing alone lifts RoboTwin +47% relative — the shared VAE (Wan-2.2 for both RGB and 3D pointmaps) plus a linear DINOv3 projection is what keeps the encoder budget flat. Real-world YAM bimanual (5 tasks, up to 20 rollouts each): full joint 83.0% avg in-distribution vs ManiFlow 58.0 / π0.5 52.1 / Fast-WAM 31.7, and OOD margin widens to 2.4× (76.1 vs 31.5). Sharpest single-task result on file: eight-stage self-repair gripper (4 mm bit into 4.5 mm socket) completes all stages in 11/20 rollouts vs 1/20 for the strongest baseline. 50%-data ablation: action-only Flex-π at half data ≈ π0.5 at full data.
• [2026-08-12] Fast-WAM: Do World Action Models Need Test-time Future Imagination?: Fast-WAM (Tsinghua IIIS + AIR) stakes out a “video-modeling representation without video-generation latency” middle position on the WAM recipe board — MoT WAM on Wan2.2-5B (6B total, 1B action expert) with video co-training during training but no future-video denoising at inference. Under shared-framework controlled comparison, Fast-WAM matches imagine-then-execute variants (Joint / IDM) within 1 pp on LIBERO (97.6) and RoboTwin (91.8) without embodied pretraining, while removing video co-training drops 4–8 pp — the video-co-training gap dominates the imagine-vs-no-imagine gap. Runs at 190 ms (>4× faster than IDM’s 810 ms). Small-scale controlled counterpart to Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2’s scaling-law result (the Slack pointer explicitly frames it as “what Dyna-2 is essentially”), and complements µ₀’s frozen-trace-WM factorization and LingBot-VA’s unified causal AR flow-matching with an explicit imagine-vs-no-imagine ablation.
• [2026-08-11] FluxVLA Engine — Standardized VLA Engineering Foundation: FluxVLA Engine (LimX Dynamics) adds a workflow / configuration framework row to the VLA infrastructure axis — a unified single-config-file substrate spanning data → training → inference → real-robot deployment, with modular interfaces for VLM/VLA/WAM backbones (Qwen, GR00T, Pi series), mainstream simulators (Isaac Sim, LIBERO, RoboCasa), and hardware (LimX Oli/TRON 2, UR, ALOHA). Distinct scope from the runtime-plane siblings already on the board (Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots llama.cpp-shaped runtime, DimOS — The Agentive Operating System for Physical Space (v0.0.13) message-passing OS): FluxVLA targets the training-to-deployment authoring loop, not the post-checkpoint deployment plane. Claims 5–10× inference speedup via operator fusion + RTC integration and a ~30-minute end-to-end workflow, but the product page carries zero benchmarks, no baseline hardware, and no per-task numbers — announcement/product-marketing surface only.
• [2026-08-11] MicroFactory — 99.9% reliability via $5 human-in-the-loop DAgger retraining on Jetson (Ilir Aliu × Igor Kulakov podcast): MicroFactory (Kulakov / Aliu podcast) stakes out the industrial-deployment side of the DAgger-not-RL thesis: precision-electronics-assembly cells targeting 99.9% reliability via human-in-the-loop DAgger corrections on Jetson Nano at ~$5/retraining cycle, explicitly rejecting RL because “RL damages hardware” in this regime — operational endorsement of the Malik sensorimotor counter-position and industrial-deployment sibling to HUD.ai’s Assemble Bench training platform. Announcement only; no per-task numbers, no released repo, YouTube transcript unavailable at filing.
• [2026-08-10] Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models: Dyna-2 (Dyna Robotics) runs the first apples-to-apples WAM-vs-VLA head-to-head under matched data + hyperparameters + three seeds — an early Dyna-2 (pre-1M, action-only) still beats production VLA Dyna-1 (Qwen3-VL-3B) with 1.55× success and 1.12× quality, winning 65% of 21 task×checkpoint cells vs 29% (6% tied) — despite the pipeline being tuned for the VLA. Deployment-side quality/throughput gap (87% vs 46%) is much larger than the in-distribution completion gap (both ~100%), evidence that world-modeling structure buys robustness that completion-only metrics miss. Sharpest filed answer to the recipe-lever debate for choosing a WAM backbone (Cosmos+GR00T, µ₀, LingBot-VA, FLUX-mimic) over end-to-end VLM+action-head recipes at scale.
• [2026-08-10] VLAff: Vision-Language-Affordance Model for Unified Actionable Affordances: VLAff (JSK Lab, U-Tokyo) extends the pointing-centric VLM lineage (Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models Embodied-R1.5, MolmoPoint) to a unified actionable-affordance output primitive — visual heatmap + 6-DoF grasp pose + trajectory in one model — trained on 204K auto-extracted human-video episodes (5.6M visual + 11.6M grasp+trajectory affordances) via SfM + hand-mesh reconstruction. Downstream execution is 3D-scene-grounded rather than a flow-matching action head; reports SOTA visual affordance prediction and zero-shot real-robot manipulation. Sits between the pointing/grounding VLM recipe (single point/bbox output) and the action-head recipes (π*0.6 / LingBot-VA), keeping the interface geometric-structural rather than continuous-actions.
• [2026-08-10] SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation: SimToolReal reformulates dexterous tool use as a single goal-pose-conditioned RL policy over procedurally generated tool primitives — 22-DoF Sharpa hand + 7-DoF KUKA arm, no per-task reward engineering, deployed zero-shot to novel tools by extracting mesh + grasp region + 6-D pose trajectory from a single RGB-D human video via SAM 3D + SAM 2 + FoundationPose.
• [2026-08-10] R&B-EnCoRe: Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning: R&B-EnCoRe treats embodied CoT reasoning as a latent variable in an importance-weighted variational inference framework so VLAs can self-supervise which reasoning primitives are action-predictive for their specific embodiment, delivering +28% manipulation / +101% legged nav / −21% AV-collision over reason-on-all-primitives baselines while halving generated reasoning tokens; validated across 1B/4B/7B/30B VLA scales.
• [2026-08-10] MEM: Multi-Scale Embodied Memory for Vision Language Action Models: MEM (Physical Intelligence + Stanford/Berkeley/MIT) equips π₀-family VLAs with dual-timescale memory — a parameter-free spatial+temporal ViT for short-horizon video memory plus a compressed language-based long-horizon memory that the high-level policy predicts autoregressively — enabling 15-minute long-horizon tasks (recipe setup, full kitchen cleanup) and unlocking in-context adaptation from failed grasp attempts in the memory context.
• [2026-08-08] χ0 (Kai0): Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies: χ0 (HKU MMLab; ICRA’26 laundry challenge team) stakes out a resource-efficient distributional-alignment recipe distinct from every scaling-axis lever on the board: three lightweight components — Model Arithmetic (weight-space merging of K subset-trained checkpoints, greedy-searched on OOD DAgger validation data), Stage Advantage (paired-observation direct advantage prediction with manually-annotated stage labels, replacing π*0.6: a VLA That Learns From Experience (RECAP) RECAP’s noisy value-difference), and Train-Deploy-Alignment (spatio-temporal augmentation + heuristic DAgger with manually-initialized failure states + temporal chunk-wise smoothing composable with RTC) — combined on a π0.5 base, hit ~250% success-rate over the strongest open-source baseline with only 20h demonstrations on 8×A100 across three garment tasks (flatten+fold, conditional retrieval, hanging) and sustain 24h non-stop live-stream operation. Direct RECAP counter-recipe on the Stage Advantage axis, and sibling to REMAC (Real-Time Robot Execution with Masked Action Chunking) / πR² (πR²: Reactive Real-time Flow Policies) on the async-inference axis with pure inference-time smoothing rather than training-time LoRA or architectural changes.
• [2026-08-08] πR²: Reactive Real-time Flow Policies: πR² (Park & Tulsiani, CMU) is the architectural + scheduling counterpart to REMAC’s training-recipe answer for async-inference on flow-matching VLAs: (1) disentangle DiT conditioning into a slow channel (vision/language features from the VLM, updated asynchronously in a background thread with a learned integer-delay embedding) and a fast channel (proprioception, refreshed every control tick), and (2) replace the standard chunk-shared noise level with a latency-adaptive staircase — clamped-clean front (in-flight actions as inpaint conditioning), linear-ramp interior emitting new clean actions per call, pure-noise tail — parameterized by measured inference delay τ, with per-position AdaLN as the only architectural change. Reports 25 Hz replanning on RTX A5000 (~4× faster than base GR00T-N1.7) via 1 NFE per call regardless of backbone size, and on four contact-rich real-world dexterous xArm6+XHand tasks beats the strongest baseline (Train-Time RTC) by up to +30 pp — largest lifts on reactivity-critical tasks (Tidy Up Book 12/20 vs 8/20, Catch Book 11/20 vs 5/20, Insert Box 16/20 vs 10/20). Diagnostic mechanism: per-tick proprioceptive refresh lets πR² modulate grip force off live fingertip contact (stops at 50 N on Tidy Up Book) while RTC’s stale plan overshoots to 120 N. Complements Real-Time Robot Execution with Masked Action Chunking (REMAC targets intra-chunk-inconsistency in the executed prefix via LoRA + prefix-mask training; πR² restructures the model + scheduler) and slots on the same async-inference axis as A2A: Action-to-Action Flow Matching (A2A: starting-distribution lever).
• [2026-08-08] AI Sapiens Open Source — ROBOTIS Docs Directory: ROBOTIS’s AI Sapiens docs page names cyclo_intelligence as the imitation-learning / policy-training component of the K1 humanoid stack, exposed via the shared DynamixelSDK + dynamixel_hardware_interface actuator-control layer.
• [2026-08-07] Shotwell releases 415+ hours of annotations on ABC-130K: Shotwell’s 415-hour annotation release on ABC-130K signals a supply-side data-enrichment layer emerging around the largest open bimanual teleop dataset — potentially plugging the language-conditioning / subtask-segmentation gap in ABC’s release.
• [2026-08-07] World-to-Wrist: Task-Conditioned Future Wrist Modeling for Fine-Grained Robot Manipulation: W²-VLA (Pan et al.) argues main-view and wrist-view play asymmetric roles and adds a task-conditioned 16-token latent interface that a wrist predictor uses to forecast future V-JEPA wrist latents, which then condition flow-matching action generation — a wrist-only, action-proximal alternative to full video world-model conditioning while sustaining >80 Hz.
• [2026-08-07] Real-Time Robot Execution with Masked Action Chunking: REMAC (Wang, Zhang, Yan et al.; Cisco Research) is the first filed training-time recipe for async-inference robustness on flow-matching VLAs — diagnoses a second failure mode alongside inter-chunk discontinuity (which BID and RTC already target): intra-chunk inconsistency, where the executed prefix of the next chunk was conditioned on a stale observation. Fixes both with LoRA (≤1.5% params) trained under prefix masking (loss restricted to the to-be-executed suffix, cut point sampled across all delays), a self-conditioned curriculum (interpolation anneals from ground-truth to stop-gradient rollouts of the pretrained policy), and a residual-alignment term (LoRA correction matched to the pretrained-vs-target flow residual), plus a prefix-preserved sampler at inference. Reports beating Naive Async / BID / RTC across all delays on 12 Kinetix tasks and three real Franka setups with widening margins at higher δ, composes additively with BID and RTC, adds zero inference latency (LoRA merges into the backbone), and extends to ACT. Slots in as a training-time recipe orthogonal to LingBot-VA’s world-model-side FDM re-grounding (Causal World Modeling for Robot Control (LingBot-VA)) and to A2A’s starting-distribution lever (A2A: Action-to-Action Flow Matching) — three different ways of exploiting the previous chunk as a structured prior for the next inference step.
• [2026-08-06] FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models: FailSafe (Lin, Duan, Fang et al.; NTU + UW + Allen AI + A*STAR) adds a fourth entry to the companion-VLM-over-frozen-VLA pattern alongside Harness VLA, LFP, and WAM-TTT — but specialized to failure states. FailSafe-VLM is queried every 10 steps by a base VLA; if a failure is detected (translation / rotation / no-ops perturbation of the current motion-planner sub-goal), it outputs an executable 7-DoF recovery delta that the robot runs directly, then control returns to the VLA. Reports +22.6 pp average lift on OpenVLA (14.7 → 37.3%), +8.0 pp on OpenVLA-OFT (90.7 → 98.7%), and +4.0 pp on π0-FAST (78.7 → 82.7%) across three ManiSkill tasks with the same checkpoint. Cross-object (+17.4 pp on Sphere / Charger) and cross-embodiment (+6.7 pp from Franka Panda-trained checkpoint to unseen xArm 6) transfer without any retraining. Distinct from Harness VLA (staging-triggered) and LFP (language-rewrite steering) in that intervention is failure-triggered on a fixed polling schedule; distinct from EgoRecovery’s gated corrective-intent module in that the recovery signal is delivered as an executable delta from a companion VLM rather than baked into the policy’s own latent space.
• [2026-08-06] RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data: RekaDaily-10k (Reka, Apache 2.0) releases 10,312 h of captioned unscripted household egocentric video — a candidate open pretraining substrate for household-scoped VLAs, with captions written against the arc of full sessions (activity + order + pauses) rather than sampled frames.
• [2026-08-06] BridgeVLA++: A Data-Efficient, Generalizable, and Memory-Augmented Vision-Language-Action Framework for 3D Manipulation: BridgeVLA++ (ByteDance Seed + CASIA + NLPR; Li, Zhu, Chen et al.) adds a 3D-aligned pipeline + in-architecture memory row to the recipe-lever board: preserves BridgeVLA v1’s core alignment trick (project point clouds → multi-view images → predict intermediate heatmaps → decode actions, keeping the pretrained VLM in its native input–output distribution) and inserts a unified spatio-temporal memory module at only ~9.2% parameter cost. Reports SOTA across five manipulation benchmarks including two memory-dependent ones while claiming to preserve v1’s data efficiency and OOD generalization, plus bimanual + additional real-world platform validation. Occupies a distinct point from RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark PrediMem’s external-keyframe-bank + predictive-coding-auxiliary and from RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR) RoboTTT’s memory-into-weights-via-TTT — three structurally different memory-augmentation answers on the board.
• [2026-08-04] FACT: Force-Aware Contact-Rich Manipulation via Timestep Modulation — demystifying when and why VLAs fail on contact-rich tasks: FACT (Parés-Morlans et al., tweet thread; underlying paper unlinked at filing) stakes out the sharpest concrete answer yet to the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position: rather than adding a new modality pipeline (Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction) or synthesizing force from vision (FELT: Generating Tactile Signals from Vision for Visuo-Tactile Manipulation), it identifies a training-side second root cause orthogonal to missing force sensing — the Beta timestep sampler used at flow-matching training starves the low-noise regime where sub-mm contact-phase action deltas live — and fixes both halves at once with a Logit-Normal post-training noise schedule (~6× more gradient in the low-noise regime, no new params/data/arch) plus time-aware force injection (current force modulates every layer, 2 s history as tokens, contact-gated learning). Reports 66% avg vs 41% for the best prior baseline across 5 tasks and ~2,500 real-world rollouts, with both components transferring across VLA backbones. Complements ModPack: An Extensible Teleoperation Interface for Bimanual Mobile Manipulation‘s data-side finding (projected joint torques give 30 pp on box-transfer contact tasks) with a training-recipe-side finding on the same underlying problem.
• [2026-08-03] Robot-eval-bench: Benchmarking Task Success Evaluation for VLMs: Kite ML’s Robot-eval-bench provides the calibration numbers the productized VLA-success-detector shape (Instance Labs — Verifying Robot Learning Episode Success) has been missing — frontier VLM judges hit 0.94 on keyframes (Gemini 3.6 Flash) and 0.92 on video (GPT-5.6 Sol) at grading real robot-policy episode success across 16 open manipulation datasets, with a general recipe (truncate/shuffle/reverse/task-swap real demos) for manufacturing failures when only success demonstrations exist.
• [2026-08-03] RL²-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models: RL²-VLA (NUS + UofT + ST Engineering; Tan et al.) stakes out a training-side + inference-side split distinct from every recipe on the board: keep the base VLA frozen, train a second lightweight offline-RL flow-matching policy conditioned on latents from the base VLA’s action expert, and at inference compose the two policies’ action-expert flow velocities rather than reranking samples or picking one head. The load-bearing empirical finding is that test-time steering follows fundamentally different scaling laws in predicted-success vs predicted-failure states — action-diversity scaling e ≈ a·k^b improves error in failure regimes and stays flat/degrades in success regimes — so RL² gates compositional steering on a conformal-prediction band over a failure-probability predictor. Reports up to +17.3 pp OOD on SIMPLER / PolaRiS and +26.7 pp on real hardware over the strongest Rephrase baseline, with an additional +18.6 pp from 8 rephrases × 5 samples + compositional steering vs single-sample inference. Sibling in framing to Learning What to Say to Your VLA: Mostly Harmless VLA Steering (LFP: frozen VLA + test-time language-rewrite steering + conformal improvement head) but steers at the velocity/latent level rather than in prompt space, and orthogonal to VINE / RECAP / QGF which fine-tune the base policy rather than composing a second one.
• [2026-08-03] Robots Acquire Manipulation Skills in Seconds from a Single Human Video (HOST): HOST (X Square Robot + Tsinghua + BIT) adds a no-parameter-update in-context skill acquisition recipe row to the lever board: cascade of task-progress estimation → predicted robot-egocentric observations → action derivation, all frozen at deployment. ~29 s per new skill from a single 3rd-person human video, 62% average success on 8-task baseline (+45 pp over strongest no-update method), and beats a 50-demo/task fine-tuned baseline at 50× fewer demos and 507× faster per skill across 50 novel tasks — the sharpest filed evidence yet that inference-time cascade prediction can substitute for demo-scaling in the narrow one-shot manipulation regime, orthogonal to the RL (π*0.6), clean-teleop (Spirit v1.5), UMI-scaling (XR-1), and unified-VLM (Embodied-R1.5) levers.
• [2026-08-03] Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand: TetherIA open-sources Aero Hand Open — $314 BOM, <400 g, 16 joints / 7 motors tendon-driven anthropomorphic hand — as a cheap embodiment substrate for dexterous-manipulation policies, including a zero-shot actuation-level MuJoCo → real RL demo.
• [2026-07-31] Gemini Robotics-ER 2 vs Claude Opus 5 clapboard eval — ER2 threw robot arms off table, hallucinated success (Robocurve / Jay Chooi): Robocurve (Jay Chooi) posts a small-N head-to-head between Gemini Robotics-ER 2 and Claude Opus 5 on a YAM-arm clapboard task via Inspect Robots — ER 2 goes 0/5 (twice hallucinating success without touching the clapboard, once flipping the robot off the table and breaking both camera mounts), Opus 5 goes 1/5 with no hallucinated successes, Gemini ~40% faster per command. First filed external third-party head-to-head on the ER model behind Gemini Robotics 2 brings whole body intelligence to robots and a sharper contact-establishment failure case than How Claude Performs on Robotics Tasks‘s LIBERO harness runs.
• [2026-07-31] Interlatent to open-source full-stack robotics deployment infrastructure — cloud inference pipelines and robot-driving stack: Interlatent (Sean / @sean_pixel) pre-announces incremental open-source release of their full-stack robotics deployment infrastructure at github.com/interlatent/interlatent — framed as both efficient cloud-inference pipelines and robot-driving plumbing, arguing no public examples of either currently exist. Third publicly-scoped entry on the deployment-runtime axis alongside Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots (model plane) and DimOS — The Agentive Operating System for Physical Space (v0.0.13) (message-passing plane), and the first filed instance targeting both halves in one platform. Announcement only — repo not readable at filing, no benchmarks, robots, or code specifics disclosed.
• [2026-07-30] OopsieData — a multi-lab dataset of real robot manipulation failures: OopsieData (Zhiyuan Zhou, 17-lab consortium) stakes out the third position on the failure-data-scarcity axis alongside EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration (egocentric human capture) and DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation (simulator synthesis): pool the real-robot failure and suboptimal rollouts that policy evals, play-data collection, and online RL already produce but throw away, across labs. Direct multi-lab federated counterpart to RW-RL Dataset: Real-World Reinforcement Learning for Robots RW-RL Dataset’s single-lab on-robot-mixed-regime bet, and the natural training substrate for the World Value Models for Robotic Manipulation Suboptimal-Value-Bench and for RECAP-style advantage-conditioning (π*0.6: a VLA That Learns From Experience (RECAP)) which require both successes and failures. Announcement only — 17 labs contributing, Oct 5 contribution deadline, CoRL 2026 workshop planned.
• [2026-07-30] Gemini Robotics 2 brings whole body intelligence to robots: Gemini Robotics 2 (Google DeepMind) is the closed-flagship successor to Gemini Robotics 1.5 and extends the ER+VLA two-tier recipe on three axes: the VLA’s action space now spans full whole-body humanoid control (Apptronik Apollo 2) plus a 22-DoF five-fingered SharpaWave hand (knot-tying, ziplock-sealing) rather than upper-body + parallel-gripper only; ER 2 sustains “several minutes and hundreds of decisions” with explicit begin/end-of-task and key-event heads for progress tracking; and — new to the series — multi-robot collaboration lets ER 2 orchestrate workflows across heterogeneous robots. On-Device 2 inherits 1.5’s “motion transfer” adaptation and reports fitting a new bi-arm embodiment (Dexmate, SO-101, Trossen) in a few hours with typically <200 examples. No headline benchmark numbers in the blog itself; safety evaluations deferred to the accompanying Safety Technical Report.
• [2026-07-30] TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM: TurboVLA (Huawei / HUST) sharpens the “skip the LLM” position sibling to Patch Policy: Efficient Embodied Control via Dense Visual Representations: reformulates V → L → A as direct V + L → A via a lightweight bidirectional V↔L interaction module between independent encoders + a compact action-chunk decoder, reporting 97.7% average LIBERO at 0.2B params, 31.2 ms latency, and 0.9 GB VRAM on a consumer RTX 4090. Two orthogonal LLM-free recipes now on file (Patch Policy: frozen ViT dense patch tokens + small transformer; TurboVLA: bidirectional V↔L interaction module) both hit LIBERO-tier results at sub-billion scale with ~10-30 ms latency, arguing the LLM-centric pathway is a design choice rather than a scaling requirement.
• [2026-07-29] HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone: HiFi-UMI runs an unusually load-bearing three-backbone data-side ablation — StarVLA-QwenPI, OpenPI-π0.5, and LingBot-VA (VLA + WAM families) — and shows that with a high-fidelity robot-free UMI corpus (3 mm accuracy, <40 μs sync, native relative pose, six-view ~200°) a policy post-trained solely on UMI data matches an in-domain teleop baseline across all three (deltas −2.5, +3.1, −0.6 pp). Adding 4,000 h of pre-training lowers 10-task action error by 41% and adds +18.1 pp real-robot success on StarVLA-QwenPI. The result is cross-backbone stable, which is what makes it a claim about the substrate rather than about a favorable interaction with one backbone’s inductive biases.
• [2026-07-29] ArmnetBench v0.1: Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farm: ArmnetBench v0.1 supplies a fixed-50-demo-budget real-hardware datapoint on the recipe-lever board across 12 tasks × 3 cells: strengthens π0.5’s overall lead (45.4% single-arm, 52.1% bimanual, ahead of π0, GR00T, ACT, Diffusion Policy, SmolVLA, MolmoAct 2), but undercuts the ‘one recipe dominates’ reading — on precise-force tasks (tool_removal, open_lamp_door) a Diffusion Policy trained from scratch beats every VLA by 50 pp margins, and cable_clip stays 0% for all seven policies. Concrete empirical support for the Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position that VLA scaling has not solved contact-establishment tasks at low demo budgets. Launch thread also reports that training flow loss and validation action error correlate with real-world success across policy families while validation flow loss does not.
• [2026-07-29] Building Worlds That Train Robots — Real-to-Sim-to-Real (R2S2R) as a Scalable Engine for Training and Evaluating Robot Policies: World Labs frames the R2S2R engine as policy- and embodiment-agnostic infrastructure: one reconstructed world serves many VLA-style policies and hardware platforms, letting sim-based checkpoint ranking gate expensive real-world evaluation for foundation-model teams.
• [2026-07-28] GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation: GigaWorld-1 is aimed at the evaluation half of the VLA pipeline: WMBench pairs real-robot teleoperation with matched policy rollouts across 324K+ paired executions to score world-model-based simulators of VLA behavior, positioning WFM-native policy evaluation on the simulated/centralized pole alongside RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies on the real-world/decentralized pole. Direct evaluation-infrastructure counterpart to RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation and Evaluating Gemini Robotics Policies in a Veo World Simulator.
• [2026-07-28] 1XWM: From Video to Action — 1X's video-pretrained world model as a NEO robot policy: 1X’s 1XWM-as-policy stakes out a distinct row on the recipe-lever board: a text-conditioned bidirectional video WM (14B, DreamGen/UniPi lineage) as the policy backbone plus a Depth-Anything + flow-matching IDM as a thin action extractor — the “video model is the policy, IDM is a thin extractor” factorization. Contrasts with LingBot-VA’s unified causal AR flow-matching (Causal World Modeling for Robot Control (LingBot-VA)) and DVA’s causal-video-as-policy (Direct Video-Action Models — Causal Video Models Are Data-Efficient Robot Policy Learners) by keeping the WM text-conditioned + off-the-shelf-shaped and pushing the recipe intelligence into the IDM design and best-of-N filtering. Also a concrete answer to the Malik sensorimotor counter-position — 1X’s follow-up moves to stereo vision explicitly to fix monocular depth undershoot/overshoot the blog itself names as the current failure mode.
• [2026-07-28] Data Pyramid for Embodied Manipulation: Data Pyramid (Ye et al., 28-author survey led by PKU / HKUST-GZ / Shanghai AI Lab / NTU) re-reads recent embodied brain, VLA, and world-action models through their data recipes rather than architectures — organizes the ecosystem as five sources (real-robot, UMI-style, egocentric + exocentric human video, simulation, general VL data) on a scalability × robot-alignment map with quality/diversity/reusability/physical-fidelity descriptors, and closes with six data-side open challenges (tactile datasets, failure/recovery data, collection pipelines, cross-embodiment action alignment, egocentric for dexterous, principled recipes). External survey counterpart to the Luma-internal Survey notes on World Action Models / VLA for robotics (index) but scoped to data composition; recipe-lever board on this page maps cleanly onto its five-source decomposition.
• [2026-07-28] ModPack: An Extensible Teleoperation Interface for Bimanual Mobile Manipulation: ModPack (Stanford REALab) validates its modular teleop substrate by training Diffusion Transformer (DiT) policies on ModPack-collected demos across two robots and two tasks. On box-transfer-with-haptic-feedback (bimanual + rack), the all-cameras + joint-torque DiT reaches 12/20 vs 6/20 for all-cameras-without-torque — a cheap datapoint for the tactile/contact open question: projected joint torques, not instrumented tactile skin, deliver the 30 pp gain, with attention-pattern analysis showing torque-token attention spikes during contact-critical grasping phases. On cloth-placement-with-active-perception, head-camera-only wins 22/25 vs all-camera 20/25 vs wrist-only 3/25 — sharpens the “which observation channel dominates” question at the DiT-head level.
• [2026-07-27] Is Diversity All You Need for Scalable Robotic Manipulation?: Shi et al. (Shanghai AI Lab / HKU / AgiBot) run the most systematic diversity-axis diagnostic filed on this page: task diversity dominates per-task quantity (episode-sampled beats task-sampled +0.1 avg at matched size), single-embodiment AgiBot-G1 pre-training beats OXE-multi-embodiment on ManiSkill / RoboTwin / real Agilex under fine-tuning (cutting against the “you need multi-embodiment data for cross-embodiment transfer” folk theorem), and expert velocity variation is a confounder rather than a benefit. Proposes GO-1-Pro’s Velocity Model — a SigLIP+MLP that predicts observation-conditional optimal chunk length used to temporally rescale demos to canonical-velocity before training — reporting +15% average gain equivalent to 2.5× pre-training data. Adds an action-representation-side lever orthogonal to π*0.6’s RL-on-flow-matching and Spirit-v1.5’s clean-teleop position: fix the temporal dimension of demonstrator variability while preserving spatial multimodality.
• [2026-07-27] TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics: TOPReward supplies a training-free dense reward substrate for the RECAP-style advantage-conditioning recipe: on a real SO-100 with only 50 noisy demos per task, TOP-AWR beats behavior cloning across all 6 tabletop tasks (10/7 doll-in-box, 10/7 pick-cube, 6.3/5.7 pen-in-cup). Contrasts with π*0.6’s binarized-text-token value function and Learning-to-Fold’s policy-as-own-Q-head by pushing the value signal entirely into a frozen video-VLM’s affirmative-token log-likelihood — no value model, no head, no fine-tune.
• [2026-07-27] RW-RL Dataset: Real-World Reinforcement Learning for Robots: RW-RL Dataset (SJTU MINT + Boden + Junpu) is the open data-substrate counterpart to Evo-RL — the first filed public dataset designed around iterated RL rather than imitation, with three coordinated regimes (teleop + human-intervention + autonomous rollout), multi-signal reward/quality/event labels per episode (rule/statistics + intervention start/end + model-assisted scores), and 12,336 bimanual episodes shipped in R1Lite (targeting 1000+ hours across 4 robot series and 9 domains). Sharpens the intervention-scaling debate by staking out a third position alongside UniIntervene: Agentic Intervention for Efficient Real-World Reinforcement Learning‘s learned-intervention-agent labor-efficiency answer and EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration‘s off-robot egocentric-recovery answer — bet directly on on-robot intervention + reward-label-completeness at scale as the missing data axis.
• [2026-07-26] Robotics Notebooks — Chinese-language robotics tech stack map and learning index: Chinese-language ‘robotics tech stack map’ — a curated learning index for VLA/robotics concepts and papers, navigated by cross-link degree.
• [2026-07-26] Action QFormer: Structured Representation Shaping under Action Supervision in Vision-Language-Action Models: Action QFormer reframes action supervision as a representation-shaping force on the inherited VLM and adds a new structural-interface lever orthogonal to the recipe/backbone/data axes already on this page: an instruction-conditioned query interface between the multimodal backbone and the action head, so action loss reorganizes information into an action-facing representation rather than broadly rewriting the shared multimodal pathway. Reports 18.8% → 56.3% average closed-loop success and 22.5% → 75.5% fixed-instruction action-generation correctness on zero-shot sim-to-real navigation, with near-elimination of OOD instruction generations — sibling in regime to Robostral Navigate: single-camera AI navigation (single-camera sim-to-real VLN from a grounding VLM) but staking out interface redesign rather than RL post-training as the lever; sibling in mechanism to Transfer between Modalities with MetaQueries‘s frozen-MLLM + learnable-queries pattern, imported from generation to action decoding.
• [2026-07-25] SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation: SuperMap (CMU, RSS 2026) supplies the persistent spatio-temporal object-identity front-end that VLN policies could sit on top of — a training-free SLAM + open-vocabulary perception fusion producing a queryable 4D scene graph with typed spatial edges and per-object trajectory history, complementing the CAD-known-pose front-end that 3D-Object Perception Transformer (3PT) provides for industrial VLAs with the open-vocabulary + persistent-identity analog for navigation.
• [2026-07-25] FELT: Generating Tactile Signals from Vision for Visuo-Tactile Manipulation: FELT proposes a latent tactile features channel for VLA policies — a frozen-encoder + lightweight-decoder pipeline that emits either a generated per-finger pressure image or its pre-pixel latent, letting the VLA ingest a contact-perception signal without any tactile sensor at training or deployment. Concrete existence proof for the “contact perception without instrumented hardware” answer to the tactile-integration question this concept page tracks.
• [2026-07-25] Claude Opus 5 driving robot arms zero-shot via Inspect Robots (Jay Chooi / Robocurve): Robocurve (Jay Chooi) posts zero-shot Claude Opus 5 rollouts on bowl-stacking and fork-on-plate real-hardware tasks and announces Inspect Robots (MIT), an open-source evaluation harness for VLA / physical-AI models, with a promised frontier-LLM-vs-VLA report — an outside-lab counterpoint to How Claude Performs on Robotics Tasks‘s finding that every LLM-guided VLA under-performs MolmoAct alone.
• [2026-07-24] A2A: Action-to-Action Flow Matching: A2A (Jia et al., RSS 2026) stakes out a new lever for fast flow-matching action heads: shorten the flow trajectory by anchoring one endpoint to the previous proprioceptive action rather than Gaussian noise, so a single Euler step (0.56 ms on RTX 5090) suffices for ≥90% task success. Complements B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations‘s output-space lever with a starting-distribution lever, and sits orthogonal to the distillation-based fast-VLA thread (RECAP / TDM-R1). Also reports the largest surviving margins on visual generalization at Level-3 perturbations (38% at 6 NFE vs FM-UNet 16% / DDPM-UNet 10%) and 80% real-world OOD success on a glowing-cube variant where FM-UNet and DDPM-UNet collapse to 0%.
• [2026-07-23] EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration: EgoRecovery (Ge et al., Fudan) treats failure recovery as a separable VLA capability: a corrective-intent module trained mostly from egocentric human recovery clips + a small robot recovery dataset, gated on by a learned recovery-state predictor over robot observations so success-mode execution stays untouched. First filed paper to attack recovery as a data-generation problem (>10× egocentric-vs-teleop rate on recovery segments) rather than as a policy-conditioning or online-intervention problem à la UniIntervene: Agentic Intervention for Efficient Real-World Reinforcement Learning.
• [2026-07-23] Towards Machines with a Thousand Hands: Generalist AI extends GEN-1 to ~9,000 end-effector variations across >500K hours of real interaction data — a new “multi-end-effector-scaling” recipe alongside the action-pretraining / clean-teleop / unified-VLM / frozen-WFM+action-expert / UMI-hour-scale levers already on the page. Introduces a task-vector diagnostic ($\Delta W = W_{\mathrm{ft}} - W_0$) decomposed across sensor-processing / harmonic-reasoning / actuation subsystems to localize which end effector stresses which part of the architecture (whisks shift sensor processing more than peelers), turning “what data to collect next” into a measurable question rather than a heuristic. Demonstrates mid-rollout end-effector swap without model updates — sharpest filed test yet of policy conditioning on embodiment within a single rollout, orthogonal to the zero-shot cross-embodiment generalization other filed entries measure across separate rollouts.
• [2026-07-23] RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning: RoboHarness (Huawei Noah’s Ark; Huang et al.) generalizes the “wrap frozen VLA in an orchestrator” recipe from a single VLA + analytic primitives (Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents) to heterogeneous learned policies (VLAs + RL + TAMP, extensible to MPC / navigation / WAMs), where the load-bearing contribution is the Memory Bridge — before invoking policy B after policy A, retrieve trajectories associated with B, estimate its in-distribution state region, and guide the robot toward that region without any joint retraining. Multi-modal execution memory + online evidence characterize each policy’s capability boundary for routing decisions. Evaluated on three public benchmarks + 500 customized tasks + 135 real-robot experiments; reports substantial gains in zero-shot long-horizon planning and OOD robustness (no per-benchmark numbers in the abstract). Companion filing to Harness VLA in the same Slack thread — the two occupy complementary points on the same “capability-aware routing over frozen policies” axis with heterogeneity as the new dimension.
• [2026-07-23] Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents: Harness VLA (RLinf; Zhang et al.) is the sharpest empirical instance yet of the “wrap frozen VLA in an orchestrator” recipe on this page — a memory-augmented LLM planner exposes π0.5 as a narrow retryable contact-rich primitive (released API: pi0_pick / pi0_doubled) alongside a fixed library of analytic primitives (grounding / staging / transport / navigation / release), and learns the operating range of each primitive from execution traces, global success rules, and failure models rather than expanding the skill library. Stricter framing than VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation (contact-only VLA invocation, not mid-rollout interruption) and than Learning What to Say to Your VLA: Mostly Harmless Vision Language Action Model Steering (broader analytic-primitive support around the frozen VLA, not just language steering). Reports +38.6 pp on LIBERO-Pro and +25.4 pp on RoboCasa365 over the strongest baselines, and 58.4% on RoboTwin C2R bimanual — all with π0.5 frozen and no VLA finetuning. The load-bearing move per the released strict_hybrid_guide prompts is staging discipline: the planner decides where the VLA should begin acting rather than asking the policy to absorb the full distribution shift itself.
• [2026-07-22] Introducing Perceptron Mk1: Perceptron Mk1 is pitched both as an offline data engine for VLA training (subtask boundaries, reward-model labels, action-conditioned annotations, episode-filtering quality scores) and as an inference-time sidecar next to a running VLA (grasp affordances, cross-camera identity, video-derived retry/progress signals) — a closed-source vendor betting that one VLM can fill both slots in the VLA stack.
• [2026-07-21] DriftWorld: Fast World Modeling through Drifting: DriftWorld (Yilun Du et al.) shows that a fast one-step drifting world model can serve as both an inference-time planner (rolling out many candidate action sequences from a base policy and selecting the highest-reward rollout) and an offline policy evaluator with r ≥ 0.92 rank correlation to ground truth on Robomimic + Push-T — deployment paths for VLA policies that don’t require faster action heads, just a faster WFM peer.
• [2026-07-21] VINE: Taming Generative Control Policies for Reinforcement Learning: VINE (AgiBot / Nanyang Tech / Sun Yat-sen; Yang et al.) adds a training-time end-to-end value-gradient RL recipe for flow-matching action heads that keeps iterative multi-step denoising, expressiveness, and BPTT-through-all-ten-denoising-steps simultaneously — the first filed instance to preserve all three. Diagnoses the instability that motivated RECAP/CFGRL, FPO, and QGF as coming from the vanilla BC sampler rather than iterative generation itself; fixes it by re-interpolating a fresh state at every denoising step at training time only, so inference remains the standard flow-matching Euler sampler. Reports beating SOTA RL on OGBench and a real-world robot manipulation task — sets up a direct head-to-head between advantage-conditioning-as-supervision (RECAP) and re-interpolated value-gradient BPTT (VINE) for improving a flow-matching action head with RL.
• [2026-07-20] Danfei Xu — two root-level paradigm shifts in robot learning: Sim2Real for locomotion, Behavior Cloning for manipulation: Danfei Xu frames Behavior Cloning as the root-level deep-learning-era paradigm shift for manipulation (paired with Sim2Real for locomotion), a taxonomy that puts the entire VLA lineage under one banner.
• [2026-07-20] Introducing Cosmos 3 Edge: Cosmos 3 Edge Policy (DROID) is a 4B on-device WAM post-trained from the shared Cosmos 3 dual-tower backbone for pick-and-place, adding an edge-hardware WFM-as-VLA-backbone row to the recipe board — competitive at 15 Hz control on Jetson Thor with 32 actions/inference at 640×360 observations, released with post-training scripts so a small H100 cluster / DGX Station can specialize it. Sibling to the flagship Cosmos 3 policy at 4B scale, and a direct contrast with clean-teleop scaling (Spirit v1.5), UMI-hour scaling (XR-1), and RECAP-style RL (π*0.6) as the lever picked for edge deployment.
• [2026-07-20] Pareto (Hebbian Robotics) indexes AllenAI's MolmoAct2 Bimanual YAM dataset — clusters + prompt normalization (Brandon Ong): Hebbian Robotics’ Pareto tool now indexes AllenAI’s MolmoAct2 Bimanual YAM dataset with cluster-discovery + prompt-normalization surfaces — a first filed instance of intra-dataset analysis tooling for VLA training data, complementing the inter-dataset catalogs (datasets.bot, TrueLabel) already on file.
• [2026-07-20] Learning to Fold: prizewinning solution at LeHome Challenge 2026 (1st place online, 2nd offline): Learning to Fold (Larchenko) is the tightest single-researcher independent reproduction yet of the π*0.6: a VLA That Learns From Experience (RECAP) RECAP recipe — a flow-matching VLA (BEHAVIOR-1K 2025 winner’s SigLIP + Gemma-2B prefix + Gemma-300M action expert) fine-tuned with AWR + RECAP-style advantage conditioning that placed 1st of 62 teams in the LeHome Challenge 2026 simulation round (79.63%, 6.1 points ahead of 2nd) and 2nd in the ICRA real-world garment-folding final. Two structural moves worth pulling out on top of RECAP: (1) policy is its own value function, Q-function, and cheap world-model substitute — one learned query token feeds cheap linear heads reading only image tokens for success, task completion, garment type, current keypoint distances, 30-frame-future keypoint distances, and an action-conditional success residual; the Q-head drives advantage estimation, live failure detection, and best-of-N candidate selection with no separate value model (a counter-recipe to the World Value Models for Robotic Manipulation “separate value model” axis and to π*0.6’s binarized text-token value); (2) AWR applied through the sampler rather than through the loss weight — high-|advantage| frames are simply loaded more often, batches stay unweighted MSE, and effective batch utilization stays at 100% of the weight mass, with auxiliary-head losses importance-weighted back to unbiased. Also introduces Thompson-sampling bandits over per-garment-type inference hyperparameters (execution length, playback speed, inpainting onset, CFG scale, noise temperature, best-of-N), asynchronous training/rollout/DAgger loop over HuggingFace Hub, and a checkpoint-rollback trick that reliably kicked the policy out of local optima. Sibling reproduction to Evo-RL: Open Real-World Offline RL on SO-101 and AgileX PiPER with different plumbing (Hub-as-message-bus vs LeRobot CLI) and a richer head stack.
• [2026-07-18] Sunday Robotics data engine — from one Memory Developer off Craigslist to 1,000+ (Perry Jia thread): Perry Jia (Data Ops, Sunday Robotics; ex-Tesla Autopilot/Optimus data-engine lead) opens a thread claiming Sunday’s ACT-2 data engine is backed by 1,000+ paid Memory Developers grown from a single Craigslist hire over two years — puts an operational-labor number on Sunday’s real-home glove-capture recipe row on the recipe-lever board, complementing the model-side Chinchilla claim from Sunday Robotics ran the Chinchilla scaling law with their VLA model (Tony Zhao follow-up) and the “embracing chaos” thesis from Low-data inductive-bias lessons don't translate to the high-data regime — embracing chaos produced emergent policy capabilities (Nishant Desai / Sunday Robotics) with the workforce-substrate that generates the chaos. No technical detail beyond the workforce-count claim; underlying thread posts (2/N…) not retrievable at filing.
• [2026-07-17] WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time: WAM-TTT (Feng, Lyu, Wang et al.) is the first filed paper form of test-time-training as a WAM steering interface — the RoboTTT tweet’s (RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR)) recipe made concrete. Stakes out a sixth structural lever: keep the WAM backbone frozen and push each demonstration into a small key–value memory updated by self-supervised video prediction on raw human videos. The load-bearing piece the tweet did not describe is a meta-training stage on paired human-robot data with a KV memory reconstruction objective that aligns human-video-derived memory content with robot-behavior-relevant features; this is what lets the human-video TTT update translate into control gains. Reports consistent wins over in-context human-video conditioning across manipulation tasks and generalization settings (no per-task numbers in the abstract). Sibling in framing to Learning What to Say to Your VLA: Mostly Harmless VLA Steering‘s LFP (frozen VLA + test-time steering via language rewrites) and to World Pilot: Steering Vision-Language-Action Models with World-Action Priors (WAM-as-policy-peer) — but the only filed instance that steers via TTT on the WAM’s own memory.
• [2026-07-17] Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation: WANDA (CMU; Guo, Li, Shi) is evaluated with π0.5 as the downstream consumer of its one-demo-to-many-trajectories synthetic data engine — reports π0.5 fine-tuned only on WANDA-generated data reaching 54.8% real-world task progress across five long-horizon Agibot G1 tasks and outperforming a π0.5 trained on 50 teleop demonstrations. Cross-embodiment story is zero-shot Linearbot deployment from an Agibot G1 source demo via world-substrate reuse + whole-body re-planning, not policy retargeting; sibling to Video2Robo and LEGS on the “3DGS-backed synthetic data as a VLA training substrate” axis, with the mobile-manipulation + contact-rich variant.
• [2026-07-17] Mixture of Frames Policy: Multi-Frame Action Denoising for Bimanual Mobile Manipulation: Mixture of Frames Policy (Dian Wang, Shuran Song, Jeannette Bohg et al.) adds a representation-side lever orthogonal to the recipe axes already on the page: instead of picking one action-coordinate frame (base / left-hand / right-hand / rel-traj), denoise the action chunk in all of them in parallel and let an MoE router per-subtask combine the experts. Reports MoF-MoE 66.8 vs vanilla Diffusion Policy 50.3 and — critically — beats the oracle-per-task frame choice (63.8) by 3 points, evidence that gains exist from within-episode frame switching that no fixed per-task frame can recover. Ablations show best-frame supervision (−6.2), auxiliary loss (−8.0), and canonical rel-traj frame (−2.5) are all load-bearing. Sibling to B-spline Policy’s structural-action-head lever, but changes the frame the action space is defined in rather than its output type.
• [2026-07-17] Sunday Robotics ran the Chinchilla scaling law with their VLA model (Tony Zhao follow-up): In-thread follow-up to Sunday Robotics’ ACT-2 Preview launch: Tony Zhao asserts they “actually ran the exact chinchilla-scaling-law” on their VLA, deferring the writeup to @nadeesha99 — no numbers or methodology in the tweet. If it lands, would be the first filed instance of Chinchilla-style isoFLOP $(N,D)$ scaling on a production home-robot VLA, sharpening Sunday’s real-home glove-capture recipe’s data-substrate-as-lever position vs teleop-scaling and UMI-scaling competitors.
• [2026-07-16] Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1): mimic robotics announces the M1 dexterous hand + U1 kinematically-locked wearable as the hardware substrate for their upcoming Video Action Model line (successor to mimic-video), taking an infrastructure-first position on VLA data recipes: teleoperation is structurally-capped by latency, retargeting friction, and operator discomfort, and only wearable-collected data at wearable-fidelity escapes those caps. The “data pyramid” (human video base → wearable middle → teleop/deployment top) is proposed as the scaling strategy, with hand morphology held constant across all three tiers to eliminate the cross-embodiment gap between pretraining substrate and deployment target — sibling to but hardware-driven rather than dataset-driven versus Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training‘s 100K-hour UMI recipe.
• [2026-07-16] RoboMemArena: A Comprehensive and Challenging Robotic Memory Benchmark: RoboMemArena (HKUST-GZ + Zhejiang + Westlake + Tsinghua) adds a memory-augmented dual-system recipe distinct from the existing recipes on this page: PrediMem pairs a Qwen3-VL-8B S2 planner with a π0.5-shaped S1 actor, an explicit keyframe bank + recent-frame buffer, and a training-only predictive-coding auxiliary head that trains S2 to anticipate next-frame ViT features (α=0.1 latent MSE + cosine) — reshaping the hidden space so keyframe-write decisions come from the standard LM head, no separate retrieval module. Reports 38.5% TSR / 55.2% CSR on their 26-task long-horizon memory benchmark vs π0.5 21.5/38.7 and MemER 27.3/49.1; ablations show removing the predictive-coding head drops TSR to 32.3% and removing the keyframe bank drops it to 17.7%. Sharpest datapoint yet against zero-shot use of frontier VLMs as memory managers — GPT-5.4 used as S2 without robotic-domain training reaches only 8.7% TSR, arguing memory-management VLMs need robotic-specific supervision rather than raw multimodal scale. Backbone-scaling curve is clean: Qwen3-1.7B 19.9% → Qwen3-4B 31.9% → Qwen3-VL-8B 38.5% TSR.
• [2026-07-16] RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination: RxBrain (Tencent) — extends a 6.2B unified embodied-cognition model to continuous robot action generation, reporting real-robot performance without large-scale action-data pretraining.
• [2026-07-16] Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training: Xiaomi-Robotics-1 (XR-1) stakes out a seventh lever — scale-of-embodiment-free-UMI-data: 100,000 hours of hand-held UMI wrist-camera pre-training across 1,700+ scenarios (VLM-auto-labeled with per-clip state-transition captions), then cross-embodiment post-training. Reports clean pre-train scaling (validation action error monotone-decreasing in data + model) that transfers to real-robot success in unseen environments after post-training. Sharpest π0.5 head-to-head yet: at <10 h/task of demonstrations XR-1 reaches 75% overall real-robot success vs π0.5’s 40% (Phone Packing, Printer Refilling, Laundry Loading, Box Packing), and <40 h/task lifts to 85%. Also SOTA on four sim benchmarks with +58.3% relative gain on RoboDojo and +23.2% on RoboCasa365. Sibling to Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model from the same Xiaomi Robotics lab — U0 augments robot data via a WFM, XR-1 replaces it with UMI scale.
• [2026-07-15] B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations: B-spline Policy (Han, Xiong, Chen, Liu, Torralba, Zhu, Du) moves orthogonally to the recipe axes catalogued here — it changes the action head’s output space from discrete action chunks to continuous B-spline curves (control points + knots as one fixed-size vector), a drop-in replacement in Diffusion Policy and ACT that composes with any backbone / pretraining / RL recipe. Because the curve is time-continuous and differentiable, the low-level controller can sample at arbitrary rate, temporally rescale to run 2–4× faster, and read velocity/acceleration off derivatives for feedforward PD; three real-world tasks show 2–4× time reduction at matched-or-better success (Cube Picking 6.48 s → 2.45 s at 20/20; Table Cleaning 41.4 s → 18.05 s at 11–15/20; DemoSpeedUp baseline collapses to 3/20 or 0/20 at the same speeds), and inference-time segment alignment is flagged as load-bearing.
• [2026-07-15] RoboTTT — 8k-timestep robot policy via test-time training (Jim Fan / NVIDIA GEAR): RoboTTT (Jim Fan / NVIDIA GEAR launch tweet) stakes out a sixth structural lever alongside action-pretraining / clean-teleop / unified-VLM-pointing / frozen-WFM+action-expert / native-video+action pretraining: memory-into-weights via test-time training. A small TTT core inside the policy takes one gradient step per observation, giving 8,000 timesteps of context (~5 min) at constant per-step inference cost — 3 orders of magnitude beyond prior robot-policy SOTA. Reports one-shot in-context learning from a single human demonstration video (unseen circuit-board assembly), mid-episode error recovery from the same TTT loop, and a Context Scaling Curve where 8K-context pretraining beats 1K by 62% with no saturation. Launch-tweet only; deep-dive report deferred.
• [2026-07-14] Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model: Xiaomi-Robotics-U0 is evaluated as an upstream data engine (not a VLA itself) for π₀.₅ — adding its style-transfer augmentation to real demos lifts π₀.₅ OOD success 36.9% → 63.2% on real-world manipulation, sharpening the case that WFM pretraining prior compounds into policy generalization.
• [2026-07-14] How Claude Performs on Robotics Tasks: Anthropic Frontier Red Team’s “Claude Plays Robotics” evaluation stress-tests the LLM-as-VLA-supervisor interface across Claude Opus 4→4.7 + Mythos Preview, GPT-5.4, and Gemini 3.1 on LIBERO manipulation and Go2/G1 locomotion — with four control interfaces (direct torques, programmatic Python, pretrained policy / MolmoAct VLA supervision, RL training). Two findings sharpen the concept: (1) every LLM-guided MolmoAct configuration on LIBERO underperforms MolmoAct running alone, but newer Claude generations (Opus 4.5/4.6/4.7) defer to the VLA less on novel LIBERO-like tasks where MolmoAct fails, and Opus 4.5/4.6 + Gemini 3.1 outperform MolmoAct alone in that regime — sharpest quantitative evidence yet for the VoLo / LFP “orchestrator over interruptible VLA” thesis, and a concrete miscalibration axis (Mythos Preview overrides more than warranted on tasks the VLA handles well); (2) direct low-level control fails on high-DoF platforms — no model stands up the 29-DoF G1 humanoid — reinforcing that VLA supervision, not direct joint control, is the realistic deployment surface for frontier LLMs.
• [2026-07-14] GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks: GaP (Berkeley/NVIDIA/CMU; Goldberg, Fan, Zhu) introduces Variational Automation as a new problem regime between fixed automation and generalist robotics — bounded workcell, non-trivial pose/geometry variation — and answers it with a Graph-as-Policy counter-recipe: a multi-agent coding harness composes a directed computation graph of typed skill nodes from a 51-skill MORSL library, then a parallel-simulation rehearsal loop autonomously localizes failures to specific nodes and edits topology + parameters until success plateaus. Across 8 new VA benchmarks, GaP holds 0.93–0.99 where π₀.₅ / MolmoAct2 / CaP-X collapse to 0.05–0.24 under pose variation, matches an expert hand-engineered bimanual graph (0.953 vs 0.987 on crate washing), and — critically for the recipe debate — stages VLAs as a wrist-camera preamble for >2× success (π₀.₅ 0.17 → 0.67 on Pack fixed). Ablations collapse to zero when the graph is removed or the multi-agent decomposition is collapsed. Same Goldberg lab as CaP-X and RATs; the sharpest structural move yet in the Code-as-Policy line — the graph is the policy, not just a way of sampling training data. Composes with VLAs rather than replacing them.
• [2026-07-14] RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks: RoboClaw (AgiBot + NUS + SJTU, Yao Mu group) stakes out a lifecycle-unification lever distinct from the recipe/backbone/data axes: a single VLM agent with structured memory drives both data collection and deployment through an MCP tool interface, keeping contextual semantics consistent across the pipeline. The signature contribution is Entangled Action Pairs (EAP) — pair every forward manipulation policy with a simpler inverse reset policy trained to restore the precondition region, so the agent alternates forward/reset and generates on-policy data without manual environment resets. Iterative EAP rollouts monotonically lift per-skill success rates on an Agibot G01 (Lipstick Insertion 2/50 → 23/50 over five 50-sample iterations, Body Lotion 21→43/50) on top of a fixed human-demo seed; the paper reports 25% absolute success improvement on long-horizon vanity-organization vs a same-data non-agent baseline and 53.7% reduction in human time investment vs manual data collection. Sibling in framing to VoLo’s “VLA-as-interruptible-tool under a VLM orchestrator” (VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation) but pushed upstream into data collection, and complementary to the agentic self-improvement axis of ENPIRE (coding-agent policy authorship) and Evo-RL (RECAP-style RL) with a distinct “agent schedules + resets fixed VLA primitives” answer.
• [2026-07-13] InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization: InternVLA-A1.5 (Shanghai AI Lab / InternRobotics) stakes out a sixth structural lever alongside action-RL / clean-teleop / unified-VLM-pointing / frozen-WFM+action-expert / native-from-scratch video+action pretraining: foresight-token bottleneck supervised by a frozen video-generation WFM. A native Qwen3.5-2B backbone keeps training on VQA + subtask prediction; a lightweight unified action expert attaches via shared full-attention while modality-specific Gated DeltaNet processing is preserved; learnable foresight tokens query the shared context for task-relevant future dynamics and are distilled against a frozen WAN 2.2-5B video model; the video branch is discarded at inference so control latency stays practical. Pretrained on 1.2M robot episodes + 3M multimodal samples, reports SOTA overall on six sim benchmarks (RoboTwin 2.0 93.2, LIBERO 98.9, LIBERO-Plus 84.8, SimplerEnv 80.8, DOMINO 27.7, EBench 35.2) and claims strongest compositional generalization on held-out instruction bindings in real-world settings — sibling in spirit to µ₀’s frozen-trace-WM + action expert and World Pilot’s dual-path WFM steering, but with the WFM interface pushed into a compact learnable-token bottleneck rather than a separate peer or trajectory abstraction.
• [2026-07-12] 3D-Object Perception Transformer (3PT): 3PT (Intrinsic) supplies the perception front-end that VLA industrial-manipulation systems need: zero-shot 6-DoF pose with sub-millimeter accuracy from RGB alone (validated on 0.57 mm-clearance DIMM insertion and 100+ cluttered sheet-metal bin-picks), no depth sensor, no per-object fine-tuning — CAD model in, pose out. Complements the world-model / policy-learning side of the VLA stack with a solved perception layer for CAD-known objects.
• [2026-07-11] LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition: LUCID (CMU) stakes out an action-data-free, VLM-free counter-recipe: no VLM backbone, no robot demonstrations, no action pretraining. Intent (short-horizon object flow + palm pose) is learned from unstructured human video with a CoTracker3-derived point-token transformer conditioned on DINOv3 patches; a sim-only PPO+distillation sensorimotor policy realizes that intent on a specific embodiment. Reports 73% average vs a Veo-3.1 open-loop-planner’s 28% on three web-supervised tasks and matching 19/30 success on push-T + cable routing across a LEAP hand and parallel-jaw gripper from 1 hr of smartphone video. Directly counters both the action-pretraining-at-scale position of π*0.6 and the unified-VLM-pointing position of Embodied-R1.5 by showing that a hand-designed intent interface plus a generalist sim policy suffices to span embodiments without any action-labeled data.
• [2026-07-10] Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio: Temporal Ratio (Mishra, Chen, Xu, Liu, Chen, Mao) names the video-action generalization gap and diagnoses it as a head–backbone coupling failure: flow-matching action heads on top of Cosmos-Predict 2.5-2B silently ignore the video prediction in both directions (misleading predictions on ID → still succeed, correct compositional predictions on OOD → still fail), so raw sweeps over adaptation strategy × video noise level × rollout horizon show no clean pattern. Introduces TR = action-attention(future) / action-attention(current) as a runtime diagnostic that collapses the noisy design-space manifold to a single future-reliance axis and varies with task phase (rises during planning, falls during precise manipulation); TR-Adaptive Guidance uses it at inference to amplify compositional video conditioning when planning and back off during manipulation, improving compositional OOD success on LIBERO spatial/goal/object and on 22 real-world bimanual YAM tasks (trained on ABC-130k) while preserving ID precision — an interpretability-first alternative to the “unify the world model and action decoder” recipe of LingBot-VA and to the “add tactile/contact modality” counter-position of the Malik / Yunzhu Li thread.
• [2026-07-10] LingBot-VA 2.0: Native Video-Action Pretraining for Generalizable Robot Control: LingBot-VA 2.0 (Robbyant / Ant Group, direct v1 successor) adds a fifth structural recipe alongside action-RL / clean-teleop / unified-VLM / frozen-WFM+action-expert: native from-scratch pretraining of the full video-action stack for causality and control, rather than adapting an off-the-shelf bidirectional video generator (v1’s approach). Instantiates it with a semantic visual-action tokenizer (ViT-AE + frozen Perception Encoder alignment + self-supervised latent-action tokenizer via IDM+FDM bottleneck), a Mixture-of-Transformers with a sparse MoE-13B-A1.9B video stream (DeepSeek-V3-style 128 experts / top-8 / auxfree balancing) and dense action expert, Multi-Chunk Prediction from Next Forcing as an auxiliary head, and Foresight Reasoning (async prediction + learned-forward-dynamics re-grounding on each freshest observation) for real-time closed-loop control at claimed 225 Hz peak. Sibling to From Foundation to Application: Improving VLA Models in Practice (LingBot-VLA 2.0) (same lab’s parallel pragmatic π-style bet on a Qwen-VL backbone) — the Robbyant lab is running both experiments in parallel; reports outperforming π0.5 and LingBot-VA v1 “by a substantial margin”.
• [2026-07-09] From Foundation to Application: Improving VLA Models in Practice (LingBot-VLA 2.0): LingBot-VLA 2.0 (Robbyant/Ant Group) triples the v1 pretraining corpus to ~60K hours (50K robot trajectories across 20 configs + 10K egocentric human video), extends the action space beyond dual-arm to include head/waist/mobile-base/dexterous-hand DoFs, and adds a two-headed future-prediction proxy task (video representation model for semantic priors + depth estimator for geometric cues). Reports GM-100 gains in a generalist setting and strong cross-embodiment long-horizon mobile manipulation across two platforms — the deployment-focused sibling to Causal World Modeling for Robot Control (LingBot-VA)‘s architectural-unification bet in the same LingBot family, and the largest robot-side open-VLA pretraining corpus filed so far.
• [2026-07-09] LIME: Learning Intent-aware Camera Motion from Egocentric Video: LIME (ETH Zürich, Sun et al.) adds a decoupled viewpoint-policy interface upstream of the VLA action space: given an RGB view + free-form intent, predict a relative SE(3) target pose via an auto-regressive observation-gain text head + continuous flow-matching pose head, trained on ~3M hindsight-VLM-labelled pairs mined from RoomTour3D + Nymeria egocentric video. Reports a single LIME viewpoint step before an unchanged VLA policy lifts LIBERO-Goal success from 26% → 74% under randomized initial pose across all 10 tasks — sharper isolation of the viewpoint-choice lever than ActiveMimic’s joint-with-manipulation pretraining, and evidence the interface composes with any downstream VLA.
• [2026-07-08] RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation: RynnWorld-4D-Policy is a world-model-conditioned VLA whose inverse-dynamics head reads the tri-modal (RGB + depth + optical flow) diffusion model’s internal representations in one forward pass, avoiding denoising-time policy latency and reporting SOTA on real-world dexterous bimanual tasks.
• [2026-07-08] Robostral Navigate: single-camera AI navigation: Robostral Navigate (Mistral) extends the “grounding-VLM-as-policy” recipe from manipulation (Embodied-R1.5) to navigation: 8B model initialized from an in-house pointing/grounding VLM, single-RGB-camera input, pointing-based action space (image-space waypoint + arrival orientation), 400K sim trajectories, prefix-cached tree-attention SFT (22× token reduction), then CISPO online RL (+3.2%); reports 76.6% success on R2R-CE validation-unseen — +9.7 over best prior single-camera baseline and +4.5 over best depth/multi-camera baseline, the first filed VLN system with published numbers behind the unified single-camera language-first framing Qwen-RobotNav announced without a report.
• [2026-07-08] Towards Practical World Model-based Reinforcement Learning for Vision-Language-Action Models: VLA-MBPO (Zhang et al., Nanjing / Mila / Yang Yu group) stakes out a UMM-as-joint-world-model recipe: finetune Bagel (a pretrained AR+Diffusion unified multimodal model) as a chunk-level world model that jointly predicts multi-view next-frames and reward tokens, then run PPO/Flow-Noise entirely inside it. Reports +9.1 avg on LIBERO (+12.2 on LIBERO-Long) over SFT with a single hyperparameter set across all suites, beats matched online RL by +3.3, and validates on 5 real-world Arx-X5 / Galaxy-R1 tasks including 3 mm cable plugging and mobile wipe-board control. Ablations: removing Interleaved View Decoding tanks wrist-view LPIPS from 0.254 to 0.454; full-horizon rollout drops LIBERO-Long from 66.8 → 52.8, empirically confirming the compounding-error blow-up their theorem predicts.
• [2026-07-06] Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots: Embodied.cpp (Xu et al., SEU-PAISys) is the first filed llama.cpp-shaped inference runtime specialized for VLAs and WAMs — portable C++ + GGUF weights + one backend abstraction across CPU/CUDA/NPU, factored around a five-layer runtime contract (input adapters, sequence builders, backbone execution, head plugins, deployment adapters) derived from an architectural survey of the VLA/WAM zoo. Ships with production servers for pi0.5, HY-VLA, and LingBot-VA today (Apache-2.0), reports 100.0%/91.0% closed-loop task success on HY-VLA/π0.5, and pairs cleanly with DimOS on the deployment surface — filling the model-plane gap between “trained VLA” and “policy running on a robot”.
• [2026-07-04] RoboWorld: Fast and Reliable Neural Simulators for Generalist Robot Policy Evaluation: RoboWorld evaluates 8 open-source VLA policies (π₀.₅, PaliGemma-Fast, and others) inside a DROID-trained neural simulator, reproducing per-policy successes and failures from a shared initial frame and matching real RoboArena rankings at r = 0.989.
• [2026-07-02] Enabling Robust Cloth Manipulation via Inference-Time Simulator-in-the-Loop Refinement: SILR (Liu et al., Fan Shi group) stakes out a sim-in-the-loop counter-recipe: keep a lightweight offline-distilled base policy, then refine every action step via prior-guided MPPI over K parallel FLASH physics rollouts seeded by an RGB→cloth-mesh real-to-sim module. Real hardware jumps from 1/10–3/10 base-policy folding success to 8/10–9/10, with component ablations showing that both the physics backend choice (FLASH beats Isaac Sim / Newton on KP error at matched K) and the RGB real-to-sim (beats point-cloud diffusion) are load-bearing — an operational answer to the sensorimotor counter-position for the deformable-object regime.
• [2026-07-01] Scale Robot Policy Evaluation with Ray (Distributed Sim-Eval on Anyscale): Anyscale’s Ray + Anyscale reference architecture for VLA policy evaluation puts concrete numbers on the serving-side constraints downstream of the recipe recipes: GR00T-N1.7-3B has 26.79 s cold-start (Ray Serve’s 30 s default health-check window is too short, must be raised to 300 s) and ~1.6 s per forward pass returning a 40-step action chunk; disaggregating the policy fleet from the Isaac Sim workers onto separate GPUs is required because both are GPU-bound, and action chunking with action_horizon = 8 lets one policy replica feed ~8× as many concurrent simulators as query-every-step would.
• [2026-07-01] DimOS — The Agentive Operating System for Physical Space (v0.0.13): DimOS v0.0.13 is the first filed deployment runtime for VLA policies — Python SDK / ROS-free stack for Unitree Go2/G1/B1 quadrupeds and humanoids, xArm/AgileX Piper arms, and MAVLink drones, organized around typed Modules (In[]/Out[] streams) + autoconnect() Blueprints + swappable transports (LCM, SHM, DDS, ROS 2), with agents refactored as first-class Modules with direct stream access; complements the recipe-side entries by owning the plane between “trained VLA” and “policy running on a robot”.
• [2026-06-30] TrueLabel — Physical AI Dataset Directory: Robotics, Humanoid & Egocentric Data: TrueLabel Physical AI Dataset Directory — third-party catalog of 658 indexed robotics / humanoid / egocentric / teleoperation / simulation dataset pages with license, modality, robot platform, and popularity metadata; sibling to datasets.bot at roughly an order of magnitude more entries, useful as a scoping dashboard when picking a VLA training substrate.
• [2026-06-30] EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data: EgoScale (NVIDIA GEAR) adds the scale-of-human-data lever to the recipe board: pretrains a flow-matching VLA on 20,854 h of action-labeled egocentric video (>20× prior work), reports a log-linear loss-vs-hours scaling law (R²=0.9983) that predicts real-robot success, and reaches +54% absolute average success over no-pretraining on five 22-DoF dexterous tasks after a small (~50h human/~4h robot) aligned-mid-training stage; one robot demo + 100 aligned human demos gives 88% on Fold Shirt, and the same prior transfers to a Unitree G1 tri-finger hand (+30%+ absolute).
• [2026-06-30] datasets.bot — Curated catalog of robotics and embodied-AI training datasets: datasets.bot — external curated index of robotics/embodied-AI training datasets (AgiBot World 2026, HIW-500, ABC-130k, GR00T X-Embodiment Sim, …) with license + modality + popularity-proxy metadata, useful as a scoping dashboard before picking a VLA training substrate.
• [2026-06-30] ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining: ACE-Ego-0 (ACE Robotics + CUHK MMLab) adds a fifth lever to the egocentric-vs-teleop pretraining axis: reliability-aware joint training across heterogeneous robot, sim, and pseudo-action-labeled egocentric streams (4.53K + 1.48K hours), with a unified action representation (camera-space actions + morphology conditioning + time-aligned chunking) and a human auxiliary loss that concentrates supervision on reliable samples — explicitly modeling per-sample pseudo-label noise rather than filtering it (HumanScale) or splitting network capacity (Li et al.); reports SOTA on RoboCasa GR1 TableTop and RoboTwin 2.0 with strong real-world bimanual transfer.
• [2026-06-30] World Pilot: Steering Vision-Language-Action Models with World-Action Priors: World Pilot (Lin et al., CASIA) adds a dual-path WFM-as-policy-peer lever to the VLA recipe board: a Latent Steering path injects a scene-evolution latent into the VLA’s perception layer, and an Action Steering path supplies an anticipated trajectory as a motion prior to the action generator — both pulled from a separately-trained World-Action Model. Reports 84.7% Total on LIBERO-Plus zero-shot OOD (SOTA at submission) with largest margins on viewpoint / geometry / deformable / pose shifts; critically, the perception-side latent prior remains effective when the supplying WAM is video-pretrained-only and never action-post-trained, decoupling the WFM training data from the policy’s action data. Sibling to µ₀’s frozen-trace WAM + action expert and ImageWAM’s KV-cache injection, but the only filed instance that steers on both perception and action sides simultaneously.
• [2026-06-30] GeneralVLA-2: Geometry-Aware Reconstruction and Governed Memory for Robot Planning: GeneralVLA-2 (AIGeeksGroup) ships two targeted upgrades to the v1 hierarchical VLA stack — GeoFuse-MV3D, a multi-view geometry-prior-guided SAM3D replacement that addresses pose hallucination on calibrated RGB-D, and a governed-memory KnowledgeBank with explicit quality/confidence/lifecycle/verifier/conflict metadata; perception evaluated on GSO-30 (small lifts vs MV-SAM3D), memory evaluated against ReasoningBank on Terminal-Bench 2.0 (+4.53% SR) and SWE-Bench Verified (+3.73% resolve), positioning governed experiential memory as a portable agent layer rather than a robot-only component.
• [2026-06-30] Segmenting Robot Video into Actionable Subtasks (WGO-Bench): Macrodata Labs ships an open subtask-annotation pipeline (WGO-Bench + Refiner) that produces the kind of segment-level labels RT-H, π0.5, and SARM consume — Gemini-3.5-Flash on contact sheets with visual-timestamp cues reaches 0.306 segmentation F1 and runs at $2.64/hr (~19× cheaper than humans) across egocentric and robot-camera videos; directly addresses the “how do we mine subtask labels at scale” question several filed VLA papers leave open.
• [2026-06-29] Flexion Reflect v1.0: The Path Towards Long-Horizon Autonomous Humanoid Work: Flexion’s Reflect v1.0 wraps a teleop-trained dexterous-manipulation VLA inside a separately-fine-tuned mission VLM (RL on top of SFT lifts 16-step mission completion 38% → 90%) and a library of RL-trained motion skills — concrete deployment evidence for the VLA-as-tool framing VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation proposed, and Flexion explicitly flags the dexterous-VLA layer as the brittlest, with plans to move it to RL.
• [2026-06-29] CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation: CHORD (NVIDIA Isaac) takes the sensorimotor counter-position one level finer than ConTrack: not just track the object trajectory, match the 6-D contact wrench (force + torque on object) the human demonstration induced. Reports 82.12% across 1,831 dexterous tasks and the cleanest filed evidence that this causal abstraction over contact transfers across kinematics — 90.77% on whole-body humanoid manipulation learned from hand-only / third-person human demos.
• [2026-06-29] SimFoundry: Modular and Automated Scene Generation for Policy Learning and Evaluation: SimFoundry evaluates 5 VLA / generalist manipulation policies (DreamZero, Gr00t N1.7, π0.5, …) on the DROID platform with sim-trained-only deployment, and reports zero-shot transfer to multi-step, articulated, and bimanual real tasks — concrete sim-to-real data point for the VLA evaluation surface beyond RoboArena’s real-world A/B comparisons.
• [2026-06-29] World Value Models for Robotic Manipulation: WVM (ByteDance Seed) is the world-model-backbone counterpart to RECAP’s VLM-backbone distributional value head (π*0.6: a VLA That Learns From Experience (RECAP)): a separate, generalist, recipe-agnostic value model designed to plug into multiple policy-extraction approaches across embodiments and mixed-quality data. Argues VLM-backbone value models structurally underfit temporal grounding + future planning that value estimation requires, and introduces hesitation RMSE on the new Suboptimal-Value-Bench as the first filed metric quantifying how well a value/reward model localizes suboptimal segments — a measurement gap previous VLA papers (Spirit-v1.5, Embodied-R1.5, LingBot-VLA) have implicitly punted on.
• [2026-06-29] In-Context World Modeling for Robotic Control: ICWM (Fudan NLP) — the arxiv paper behind the HuggingPapers tweet 2070794187063099893 — In-Context World Modeling for Robotic Control (ICWM) stub, now resolvable. Frames camera/morphology generalization as in-context system identification: short self-generated, task-agnostic interaction prefix is fed into the policy context window so the model implicitly infers world dynamics before task execution, no parameter updates required. Structurally distinct from the four existing levers on this page (action-pretraining, clean-teleop, unified-VLM pointing, frozen-WFM+action-expert) — puts the generalization lever in the test-time context window rather than in pretraining; reports significant gains over standard VLA baselines on novel camera viewpoints in sim and on real robots (no numbers in the abstract).
• [2026-06-29] Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack: Tencent Hunyuan’s HyVLA-0.5 technical report — third major industrial VLA stack (after Qwen-Robot Suite and π*0.6) spanning data collection, model design, continued pre-training + SFT, RL post-training, and real-world deployment, joining the action-pretraining-plus-RL recipe family; only abstract retrievable at filing time.
• [2026-06-29] PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation: PhysisForcing shows that physically-aligned video world models yield stronger representations for downstream policies, raising closed-loop success on the WorldArena action-planner protocol from 16% → 24% and improving downstream VLA policy success.
• [2026-06-28] HuggingPapers tweet 2070794187063099893 — In-Context World Modeling for Robotic Control (ICWM): HuggingPapers tweet stub for “In-Context World Modeling for Robotic Control” (ICWM) — claims zero-shot adaptation to unseen cameras and new morphologies via in-context conditioning rather than pretraining scale or embodiment-invariant interface; underlying paper not linked from the tweet and not locatable at filing.
• [2026-06-28] UMI-Bench 1.0: An Open and Reproducible Real-World Benchmark for Tabletop Robotic Manipulation with UMI Data: UMI-Bench 1.0 — first real-world benchmark dedicated to UMI-style wrist-camera manipulation policies, with a local-first auditable protocol and seen/unseen splits across 9 tabletop tasks; complements RoboArena’s distributed crowd-sourced approach.
• [2026-06-26] Survey notes on World Action Models / VLA for robotics (index): Luma-internal survey-notes index (Dong Guo + co-author) that organizes the WAM/VLA space by directions (e.g. video-as-policy and policy-first / data-flywheel) with per-direction capability-boundary commentary — the strongest internal synthesis filed so far, complementing the public Reuss WAM taxonomy (Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models) and the World Labs functional taxonomy (A Functional Taxonomy of World Models).
• [2026-06-26] Evo-RL: Open Real-World Offline RL on SO-101 and AgileX PiPER: Evo-RL (SJTU + Evo-Tech) is the first filed open-source community implementation of the π*0.6 RECAP recipe on commodity hardware — SO-101 and AgileX PiPER — with --value.type=pistar06 as the default value function and an advantage-conditioned indicator injected as text into the policy task prefix; ports the closed-loop iteration directly into the LeRobot CLI and pairs it with the open RW-RL dataset.
• [2026-06-26] ACT in the browser — three.js SO-101 + leader-arm teleop + trained model running locally on the web: @pham_blnh demos the full lerobot ACT loop running in a browser — three.js SO-101 scene + real leader-arm teleop + ACT training + in-browser inference at lerobot.binhph.am — extending the mjswan: real-time interactive MuJoCo + ONNX policy simulation in the browser “policy in the browser” pattern to cover data collection and training, not just deployment.
• [2026-06-25] ZeroDex: Zero-Shot Long-Horizon Dexterous Manipulation via Multi-View 3D-Grounded VLM Reasoning: ZeroDex (SNU) stakes out a fifth lever beyond action-pretraining, clean-teleop, unified-VLM-pointing, and frozen-trace-WM+action-expert: a zero-shot training-free recipe where a frozen VLM produces reference-view 2D keypoints, multi-view fusion (triangulation + reference-view ray voting) lifts them to 3D, and an atomic-action library handles execution. Reports beating fine-tuned VLA baselines on novel scenes for both pick-and-place and tool-use, with closed-loop replan for long-horizon tasks.
• [2026-06-25] RAIDEN: A Toolkit for Policy Learning with YAM Bimanual Robot Arms: RAIDEN (TRI) is the first standalone MIT-licensed end-to-end VLA data-collection toolkit on file — calibration, leader-follower / SpaceMouse teleop, heterogeneous-camera mixing (ZED + RealSense), four depth backends incl. Fast Foundation Stereo, manipulability-aware IK via PyRoki/J-Parse, and WebDataset shard output; sits upstream of every VLA training recipe tracked here and is the standalone version of the data-pipeline plumbing ABC released as part of its own stack.
• [2026-06-24] VLA에게 부족한 결정적 감각 | 로봇에게 Force·Tactile이 반드시 필요한 이유 (Why VLAs Need Force and Tactile Sensing for Robot Manipulation): First filed survey of the VLA + force/tactile design space — a Korean robotics-engineering channel explainer that groups recent attempts into three architectural patterns (force injected into the model, tactile deeply fused with vision/language, VLA wrapped with tactile feedback around planning/control) plus a touch-as-language framing where touch conveys softness/hardness/slip/pressure/misalignment that aren’t visible in RGB. Useful taxonomy crosscut for the open question on tactile/contact integration that the page already lists.
• [2026-06-24] Next Forcing: Causal World Modeling with Multi-Chunk Prediction: Next Forcing trains a causal-AR visual-dynamics core for VLAs using multi-chunk prediction heads (MTP transferred from LLMs); benchmarks directly against LingBot-VA on RoboTwin and reports +93.1% relative gain at 5k steps with 2.3× faster convergence at 50 fps.
• [2026-06-23] ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?: ImageWAM (SJTU + EIT) argues video generation is the wrong prior for WAMs and replaces it with pretrained image editing consumed via KV cache; flow-matching action expert reads the editing denoiser’s caches rather than a decoded target frame, claiming to beat VLA baselines and competitive WAMs at ~1/6 FLOPs and ~1/4 latency without policy pretraining — a fifth WAM/VLA recipe alongside video-rollout-as-policy (DVA), frozen trace-WM + small action expert (µ₀), unified causal flow-matching (LingBot-VA), and clean-teleop scaling (Spirit v1.5).
• [2026-06-23] GPS: Geometric Primary Structure for Articulated Parts Perception in Robot Manipulation: GPS (SJTU, CVPR 2026 Findings) is the strongest evidence yet for a perception-rich + heuristic-policy counter-recipe to end-to-end VLAs: trains a single PointNet++ predictor of a 3-keypoint articulated-axis abstraction on 41K VR-annotated frames, then uses it to re-score AnyGrasp proposals and roll axis-constrained waypoints — 73% real-robot success across 270 trials on 9 unseen objects with no in-domain fine-tuning, vs 33% for pose-based CAPNet and 35% for flow-based GFlow.
• [2026-06-23] Causal World Modeling for Robot Control (LingBot-VA): LingBot-VA (Robbyant/Ant Group, third LingBot paper) unifies the world-model and action-decoder of the prior LingBot-VLA and LingBot-World releases into a single causal autoregressive flow-matching model over interleaved video+action tokens, processed by a dual-stream MoT (Wan2.2-5B video stream + asymmetric narrower action stream). Two transferable contributions: Noisy History Augmentation (training the action decoder on partially noisy video latents lets inference skip half the denoising steps) and an FDM-grounded asynchronous pipeline (refreshes cached forecasts against the latest real observation before predicting the next action chunk, fixing the open-loop drift that naïve async exhibits). Counter-recipe to π*0.6’s action-pretraining + Embodied-R1.5’s pointing-centric VLM by leaning on unified world+action modeling rather than separate components.
• [2026-06-22] Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI: NVIDIA Halos for Robotics — the first filed wiki artifact addressing the deployment-time substrate (IGX Thor + IEC 61508 SIL 3 Functional Safety Island + Halos OS) that VLAs would have to integrate with for industrial certification; Outside-In Safety Blueprint’s OOD-detecting Safety AI Monitor with mandatory fall-back to onboard safety is a concrete production answer to VLM/VLA perception-failure modes that purely capability-centric papers leave open
• [2026-06-21] HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos: HumanEgo (UMD) proposes a fourth lever in the egocentric-vs-teleop debate alongside HumanScale (filtering), ActiveMimic (viewpoint action), and Li et al. (per-embodiment specialization): entity-relative Interaction-Centric Tokens plus three free auxiliary forecasting heads (3D object motion, 2D image-plane trace, latent consistency) that act as an in-policy world model — and is the only filed recipe to claim outperforming matched-time teleop (+41% absolute) at 15–30 minutes of human video per task.
• [2026-06-20] Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction: Tactile-Reactive Dexterous Hand directly addresses the open “how to incorporate tactile/contact dynamics into the VLA stack” question — proposes variable-rate Mix-of-Transformer + temporal tactile VQ-VAE encoder to fuse high-frequency touch with vision/language without losing existing VLA capabilities, with 100h of motor-primitive-prioritized teleop data; >30% absolute success-rate gain on 12 delicate-force / deformable-object tasks. Most concrete answer yet to the Malik / Yunzhu-Li sensorimotor counter-position thread.
• [2026-06-19] EgoInfinity: A Web-Scale 4D Hand-Object Interaction Data Engine for Any-View Robot Retargeting and Video-to-Action Robot Learning: EgoInfinity (Hang group, Rice) — third paper in a single week (with DO AS I DO and ConTrack) arguing the human-video-to-robot pipeline is a credible alternative to teleop-data scaling; distinct framing as a modular engine whose value compounds as upstream perception models improve, with any-view retargeting that tolerates partial body visibility.
• [2026-06-19] What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?: Li et al. (MIT/UIUC/Improbable AI) stake out a third lever in the egocentric-vs-teleop debate alongside HumanScale (filtering) and ActiveMimic (viewpoint-action): per-embodiment specialization in the vision encoder and policy network, plus 3D hand-pose label quality from multi-view triangulation. Releases TriHands (532 cooking videos / 28h, triangulated 3D hands) and reports +29.7% absolute success in the low-robot-data cotraining regime — and finds that accurate hands alone are insufficient if the motion gap isn’t absorbed by per-embodiment capacity.
• [2026-06-19] Learning What to Say to Your VLA: Mostly Harmless VLA Steering: Language Feedback Policy (Jeong/Swamy/Bajcsy, CMU CoRL submission) — leaves the VLA frozen and steers it through test-time language rewrites discovered by interactive search, then distilled into an LFP with a conformalized improvement head that refuses OOD interventions; reports +24.7% sim / +65.0% Franka hardware over base π0.5 and matches VLA-SFT at ~1/5 the data, a black-box-VLA counter-recipe to the action-pretraining / unified-VLM / clean-teleop positions already on the page.
• [2026-06-19] ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control: ConTrack (UCSD; Xiaolong Wang group, ECCV 2026) attacks the kinematic-gap problem at the sensorimotor level the Malik counter-position calls out: treats object tracking as a hard RL constraint and uses a dual-variable update to reallocate authority between object fidelity and hand-style fidelity online, eliminating per-clip reward tuning. Complements Do as I Do: Dexterous Manipulation Data from Everyday Human Videos (kinematic reconstruction + retargeting) as the closed-loop-control half of the same dexterous human-video → robot pipeline.
• [2026-06-19] HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining: HumanScale supplies the sharpest quantitative datapoint to date in the egocentric-vs-teleop pretraining debate: at matched data volume, egocentric pretraining beats teleop by 24% on validation loss, 52.5% on in-distribution real-robot success, and 90% on OOD success — the OOD>ID gap directly supports the “diversity is the lever” interpretation that HumanNet and EgoVerse argued qualitatively.
• [2026-06-19] MolmoMotion: Language-guided 3D motion forecasting: MolmoMotion proposes a third pretraining target for VLA backbones — language-conditioned 3D motion forecasting rather than action prediction (π*0.6) or pointing/grounding (Embodied-R1.5) — and reports ~20-point gains in MolmoBot pick-and-place success (76.3 vs 56.0) and 5–6× sample efficiency when the same downstream policy is initialized from MolmoMotion vs. the Molmo 2 backbone it was built on.
• [2026-06-18] PressureVision++: Estimating Fingertip Pressure from Diverse RGB Images: PressureVision++ — vision-only fingertip pressure estimation supervised by prompted contact labels rather than instrumented surfaces; a concrete (if constrained) attempt at the tactile/contact perception axis that the Malik counter-position argues VLAs under-weight, surfaced in a Slack thread questioning whether force/torque/contact inputs are scalable enough to plug into a VLA.
• [2026-06-18] ActiveMimic: Egocentric Video Pretraining with Active Perception: ActiveMimic argues the human-video-to-robot gap is a signal-design problem rather than a scale problem: recovering synchronized camera + wrist trajectories from a single body-worn RGB stream and treating camera motion as a learned viewpoint action during egocentric pretraining is reported to match robot-data-pretrained baselines on real-world tasks — a sharper outcome than concurrent egocentric data papers’ qualitative “more data helps” findings.
• [2026-06-17] ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation: ABC ships matched architectural ablations across the two dominant VLA design points — a 2B DiT with tiny vision encoder (1.93B head / 85.7M DINOv3) and a 4.3B Gemma-3 VLM with a 45M action head — and reports that DiT-head scaling wins per-FLOP, while the lopsided VLM/head split lets one VLM forward be amortized across k diffusion draws. Counter-recipe to π*0.6’s RL-on-flow-matching and Embodied-R1.5’s pointing-centric VLM, and a direct sibling to Spirit v1.5’s Qwen3-VL+DiT recipe with cleaner scaling evidence.
• [2026-06-15] μ₀: A Scalable 3D Interaction-Trace World Model: µ₀ argues VLAs should be factored as a frozen action-free 3D-trace world model + a small embodiment-specific action expert, not an end-to-end VLM-with-action-head. The headline claim is that trace-conditioned policies built this way are competitive with π₀ on downstream manipulation despite µ₀ never seeing an action label — a structural counter-recipe to both the action-pretraining position (π*0.6) and the unified-VLM position (Embodied-R1.5).
• [2026-06-15] VoLo: A Physical Orchestrator for Open-Vocabulary Long-Horizon Manipulation: VoLo reframes the VLA-design debate by wrapping a VLA/WAM as an interruptible tool under a VLM orchestrator (VoLoAgent) that monitors and steers mid-rollout, rather than treating the VLA as a frozen end-to-end policy — and reports substantial gains over single-VLA and single-VLM systems on a new long-horizon open-vocabulary benchmark (RoboVoLo), with real-robot validation.
• [2026-06-14] LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting World: LEGS demonstrates that VLAs (π0, π0.5, GR00T N1.6) fine-tuned purely on 3DGS-backed synthetic data match or exceed human teleoperation across 1,110 real Unitree G1 trials — a data-side counter to Spirit-v1.5’s “clean teleop is the lever” position on the same backbone family.
• [2026-06-11] Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models: Unified 8B Embodied Foundation Model + VLA — argues that strong VLM-level pointing/grounding (Planner-Grounder-Corrector loop on a single model) substitutes for action pretraining; reports beating π0/π0.5/SpatialVLA/OpenVLA across SimplerEnv suites and matching the best-action-pretrained baselines on LIBERO.
• [2026-06-06] Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level: Counter-position arguing the VLA framing under-weights tactile / contact / sensorimotor dynamics and that CV researchers entering robotics shouldn’t skip the sensorimotor level.
• [2026-05-25] π*0.6: a VLA That Learns From Experience (RECAP): π*0.6 / RECAP — first filed flow-matching VLA trained end-to-end with offline RL on real-robot data via CFGRL-style advantage conditioning.

Open questions

• Does the unified-VLM PGC recipe (Embodied-R1.5) actually scale to the long-horizon real-world deployment regimes π*0.6 targets (13-hour espresso sessions, factory packaging) or does it remain a benchmark phenomenon?
• Is the SimplerEnv→LIBERO performance of action-pretraining-free VLAs a data-distribution artifact (both suites being heavily sampled in the embodied-VLM training corpus) or a robust generalization signal?
• How should tactile / contact dynamics be incorporated into the VLA stack — as an extra modality conditioning the VLM, as a separate low-level controller below the VLA, or via simulation-augmented training?
• For open VLAs built on Qwen-VL: how much of the gap to closed flagships (Gemini Robotics 1.5) is from the backbone choice vs the policy-training recipe vs the data?
• Does advantage conditioning generalize beyond flow-matching action heads (e.g. to FAST-tokenized discrete-action VLAs), or is it specifically a workaround for the no-log-likelihood property of flow models?

Papers

2026-06

• Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models — Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models (published 2026-06-11)

2026-05

• π*0.6: a VLA That Learns From Experience (RECAP) — π*0.6: a VLA That Learns From Experience (RECAP) (published 2026-05-25)