Human-to-Robot Retargeting

TL;DR

Human-to-robot retargeting is the problem of closing the kinematic gap between cheap, abundant human demonstration sources (mocap, egocentric RGB, third-person video) and target robot embodiments whose morphology + contact mechanics differ from a human. The filed papers disagree on where the right interface lives: physics-in-the-loop bilevel optimization over a parameterized retargeting (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting), kinematic 4D hand-object reconstruction + retargeting (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), dual-variable RL between object-tracking constraint and motion fidelity (ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control), or dropping the unreliable channels entirely and using a low-DoF embodiment-shared action signal — 3D wrist translation in the head-camera frame — with interleaved-action-token training that absorbs heterogeneous data (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots).

Key claims

• Treating humans as another bi-manual 6-DoF embodiment is structurally wrong: hand-pose estimates from monocular RGB are noisy, and a human wrist’s rotation is semantically misaligned with parallel-gripper rotation because contact patterns differ (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots §Abstract). A 3-DoF wrist-translation-only bridging signal beats the standard 6-DoF baseline by +9.91 absolute task progress, +10.0 absolute success on 15 ByteMini bi-manual tasks trained from scratch (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots Results table).
• Putting the physics simulator inside the outer retargeting loop — not just at evaluation time — produces references that downstream imitation/RL can actually track on hardware, eliminating foot sliding, self-collision, and dynamically infeasible trajectories that plague kinematic-only retargeters (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting §Abstract). Validated on two humanoids + a quadruped from a human source.
• Monocular RGB-only human videos — egocentric or third-person, in-the-wild — can be converted into robot-complete dexterous manipulation trajectories by a single pipeline that does 4D hand-object reconstruction followed by kinematic retargeting onto a multi-fingered robotic hand (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos §Abstract). Reports SOTA on both hand-object-interaction estimation and dexterous trajectory extraction.
• Object-tracking can be reframed as a hard RL constraint rather than a weighted reward term: a dual-variable update reallocates authority between object fidelity and hand-style fidelity online, eliminating per-clip reward tuning (ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control §Abstract).
• Heterogeneous data sources with different action-component coverage (in-the-wild human → only wrist; in-lab human → wrist + gripper; teleop → wrist + 6-DoF EE + gripper) can be jointly trained in a single VLA via interleaved action tokens + attention masking that limits supervision to the available components per sample (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots §Method, data table).

Recent contributions

• [2026-11-06] Fast and Accurate Hand Estimation in Pixels, Interaction and World Space: FAH (Cao) releases a second real-time monocular hand foundation model open-source alongside EgoHand v1.0 — NVIDIA monocular 3D hand pose + MHR/SOMA foundation model for robotics EgoHand v1.0 — but takes the opposite recipe (fit MANO/MHR to predicted keypoints + surface points rather than regressing parameters), bundles optional contact estimation without requiring object geometry at inference and world-space camera placement into the same model, and ships ViT-S/ViT-L sizes spanning 125 fps (no fit) to 65 fps (with mesh fit) on an RTX 4070 Ti SUPER. Candidate drop-in upstream perception stack for the kinematic-reconstruction branch (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos Do as I Do) and the object-side contact-interface branch (C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video C2Dex) where prior stacks either didn’t emit contact + world-frame camera together or didn’t hit the 30+ Hz budget.
• [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-simultaneously cell to the interface-location axis: bridge the embodiment gap at both the observation side (geometry-anchored alignment grounds human and robot observations in a compatible representation) and the action side (decoupled-decoder tokenizer shares an action latent space without forcing both domains through a single action mapping), as a co-training recipe over unpaired human and robot demonstrations against a pretrained VLA. Distinct from the action-interface-only recipes (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots reduce-to-3-DoF-wrist, LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID task-intent, World Action Learning via Interaction-Centric Spectral Latent Guidance WING frequency-domain latents) by attacking the observation + action gaps jointly, and from the representation-centric cells (CausGround’s factorized causal preconditions/effects, FACT’s force-aware human pretraining) by keeping the retargeting at the policy-adaptation layer rather than at planning or reward. Reports real-world object transfer 8.0 → 51.8, skill transfer 62.7 in settings where co-train baselines totally fail to transfer human-demonstrated skills, and out-of-distribution object-placement shift 6.0 → 45.3. Code (Apache 2.0) and project page released; built on the RLDX-1 backbone.
• [2026-11-04] EgoHand v1.0 — NVIDIA monocular 3D hand pose + MHR/SOMA foundation model for robotics: EgoHand v1.0 (NVIDIA) releases the upstream-stage substrate that recipes on this page implicitly depend on: an 883.5M DINOv3+TransformerDecoder model that turns a single 256×256 RGB hand crop + intrinsics into SOMA pose + 27D MHR shape/scale + full-perspective camera translation. Open-weights (OpenMDW 1.1 + Apache 2.0) counterpart to the closed in-house pose pipelines in Natura-1 (claimed 41% MPJPE reduction vs open baselines), EgoScale (22-DoF joint retargeting at 20,854 h), and Dyna-2 (wrist-only at 1M+ h) — makes the “convert human video into hand supervision” step reproducible outside well-funded labs. Does not stake a new interface-location position itself; it’s infrastructure for the retargeting-at-collection-time branch rather than a recipe. Inference pipeline (NVIDIA/IsaacCapture) and finetuning stack (NVlabs/GEM-X) shipped alongside the weights.
• [2026-11-04] AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data: AIRoA-0 (AIRoA; Wu, Ota, Li, Yuasa et al.) pushes the “make-the-collection-tool-embodiment-invariant” branch to its no-robot-data extreme: π0.5-shape VLA pre-trained on 16,151 hours of bimanual YUBI handheld-gripper demonstrations with no teleoperation, RL, or on-robot corrections anywhere in the pipeline. Reports 78% average success on ten bimanual Franka tasks at a never-recorded-at site vs ≤27% for four other VLAs post-trained on the same per-task handheld demonstrations, with single-checkpoint deployment across bimanual Franka, TX-G2 semi-humanoid, OpenArm, and UR via gripper swap. Operating point between Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1 (100K-h UMI + 7,200 h teleop post-train) and HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone HiFi-UMI (fidelity-first at 2K-h) — sharpest filed evidence that 16K-h of handheld data at adequate fidelity can skip the teleop post-training leg entirely on bimanual tabletop tasks at the Franka scale. Scaling study (6.25% → 100% pre-training data) reports power-law action-loss scaling that post-training does not compensate for. Load-bearing architectural pieces: visual-evidence-budget + two-level language at pre-training, task-stage prediction + 4-token 10-step gripper-history memory at post-training. Dataset, weights, and post-training code pledged.
• [2026-11-04] Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation from Sensorized Human Demonstrations: Pixel-UMI (Liu, Tao, Pathak, Salakhutdinov; CMU) stakes a scene-aware whole-body retargeting cell on the interface-location axis distinct from WARP: Whole-body Retargeting from Offline Human Motion via Closed-Form c-SEW WARP’s closed-form c-SEW IK (fixes IK, no scene conditioning), ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting ReActor (physics-in-outer-loop, no scene reconstruction), and Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s reduce-the-action-channel: Environment-Aware Retargeting reconstructs scene geometry directly from per-demonstration depth and uses it with head + gripper trajectories to produce collision-aware whole-body joint and base targets, plus delay-calibrated action assignment and Retargeting-Error-Aware Masking that probabilistically drops views with large kinematic residuals. Two rounds of 60 DAgger-corrected rollouts close the remaining gap on contact-rich long-horizon tasks. Reports 87.0% average task progress across articulated-object manipulation, deformable-object handling, constrained-space reaching, whole-body collision avoidance, and contact-rich bimanual coordination — first filed instance pushing scene-aware retargeting to whole-body mobile at this fidelity.
• [2026-11-04] Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation: Pixel (Tao, Liu, Salakhutdinov, Pathak; CMU) extends the hardware-first “make the collection tool embodiment-invariant” branch (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic, 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, Getting a Grip on Robotic Data Collection — Co-Designed Koala Gripper for Force-Mediated Manipulation RAI Koala) from parallel-gripper/single-hand regimes up to a wheeled-humanoid with whole-body control, co-designing a <$15K wheeled humanoid with a <$1K sensorized human data-collection system so camera viewpoints, gripper geometry, and task-relevant motion align between human demonstrator and robot. Companion to Pixel-UMI: Learning Environment-Aware Whole-Body Manipulation from Sensorized Human Demonstrations which carries the policy-learning recipe.
• [2026-11-02] 3D World Simulator: Hybrid Particle-Voxel Dynamics for Multimodal 3D World Simulation: 3D World Simulator factors the robot embodiment out of the generative model: arms are a separate 3DGS asset augmented with a kinematic model driven by predicted joint states, so one predicted box episode is re-drawn with the original dual arms and then with a pair of SO-101 arms across four scanned rooms (§5) without touching the dynamics model. This is a post-hoc embodiment-swap mechanism rather than a demonstration-level retargeting pipeline, but it adds an “WFM-level embodiment interchange” datapoint to the cross-embodiment axis.
• [2026-11-02] WUJI Hand 2 autonomous pen spinning with disturbance rejection via RL + motion-reference tracking: WUJI Tech’s own LinkedIn demo of WUJI Hand 2 (20-DoF) performing autonomous pen spinning with external-disturbance rejection, trained via sim RL + motion-reference tracking — the vendor-internal counterpart to the Dex-SuperMimic — Reproducing SuperMimic on Dexterous Hand Manipulation Dex-SuperMimic community reimplementation on identical hardware with an identical recipe shape (motion-reference + residual control over sim RL). Reinforces that motion-reference tracking + sim-to-real is converging as the default recipe for in-hand pen-spin on commercial anthropomorphic hands across vendor-internal and community implementations. Marketing-surface only — no quantitative disturbance-rejection result, no sim-to-real gap measurement.
• [2026-10-30] Kalyan — dense Argus-format annotations for >5,000 hours of public robot data, labeled by GPT-6.1 Sol: Kalyan’s Argus-annotation release ships dense per-hand subgoals + action labels + object-state-change timelines over DROID, MolmoAct2 YAM, ABC-130k, GM-100, and ABC-Sim — the core teleop + UMI substrates this concept tracks — now with VLM-authored structural segmentation that lets downstream policies target per-hand atomic skills instead of whole-episode action chunks. Sibling data-enrichment to Shotwell releases 415+ hours of annotations on ABC-130K Shotwell at ~12× the hours, built on the Argus harness (Argus: An Open-Source Annotator for Robotics Data).
• [2026-10-29] AdaDex: Learning Adaptable Dexterous Skills via Hybrid Dexterous Mimicking: AdaDex (Kuang et al.) stakes a hybrid-environment-population position on the interface-location axis: keep the full dexterous channel + single human reference trajectory, but train one shared policy jointly on two populations of massively parallel environments — motion-tracking (dense reference-matching reward, instills human prior) and generalization (noised initial states/goals, sparse object-success reward, forces off-reference adaptability). Sibling to REGRIND: A Minimalist Retargeting-Guided RL Recipe for Dexterous Manipulation REGRIND on the “single human demo → RL dexterous skill” axis but attacks the problem via reward-population mixture rather than retargeting-as-prior-only; sibling to ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control ConTrack on dual-objective RL over human motion but splits authority across environments rather than via a dual-variable update inside one reward. On GRAB (12 sequences, 36 benchmark motions), one-motion-per-object training reaches 96.74% progress on seen motions, 80.02% on similar-but-unseen motions, and 51.80% / 27.08% SR on random goals — vs 35.20% / 7.28% / 1.72% for ManipTrans (motion-tracking baseline) and 9.72% / 13.68% / 44.60% for SimToolReal (goal-reaching baseline). The w/o GenEnv ablation (97.47% seen → 31.75% unseen → 14.53% progress / 0.52% SR on random) is the sharpest filed isolation of generalization-environment RL as a distinct lever from motion-tracking RL for off-reference dexterous adaptability. Transfer: zero-shot Isaac Gym → MuJoCo closed-loop on all 36 motions, four hand embodiments (ArtiMANO/XHand/Sharpa/Wuji) via the same method, and real-world XHand+xArm rollouts at 30 Hz. The accompanying HOI reconstruction pipeline over casual RGB-D phone video reports better quality + lower runtime than Do-As-I-Do and EgoInfinity, making “seconds-of-human-motion-per-skill” the data-side counterpoint to the scale-side bets of EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale (20K+ h) and The Will to Act — Natura-1 Natura-1 (1M h).
• [2026-10-29] Dex-SuperMimic — Reproducing SuperMimic on Dexterous Hand Manipulation: Dex-SuperMimic (CumulusAlpha) is a third-party reproduction of SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model SuperMimic that extends the no-explicit-retargeting-in-loop recipe from humanoid whole-body control to dexterous in-hand manipulation on the Wuji Hand 1/2 and pen-spin task — datapoint that the differentiable-WFM + BPTT recipe is portable from humanoid loco to dexterous manipulation by small groups chaining community reimplementations (supermimic_vibed → Wuji MJLab → InstinctMJ), with task-awareness living in reward shaping (cage/contact rewards + absolute EMA control for reorient, motion-reference + residual control for pen spin) rather than the training loop.
• [2026-10-28] CARD: Contact-Anchored Retargeting for Dexterous Manipulation: CARD (Dai, He, Hogan et al.; Amazon FAR internship + CMU/Berkeley) stakes a contact-force-targets-as-shared-interface position on the interface-location axis: one contact anchor per hand region unifies monocular video (via the Vid-HACO video-to-contact model that beats image-HACO F1 63.3→67.2 / AP 54.3→72.2 on HRDexDB), tactile gloves, and motion capture under one representation, and the retargeter emits explicit per-anchor force targets alongside kinematics so an integral-dominant tactile admittance tracker can realize them on hardware open-loop. Reports contact preservation 43–78% vs <10% (kinematic) / 22.5% (object-aware) across video+glove+mocap+GRAB, and lifts physics-based retargeting on GRAB from 12%→32% strict task success when CARD’s contact objectives are grafted on. Sibling to CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation CHORD on the contact-wrench-as-interface axis but with the force channel materialized as controller set-points for sim-free open-loop replay rather than as a policy-reward signal; to C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video C2Dex on the object-side contact axis but anchored on hand regions rather than canonical object space; and to EgoCPT: Learning Dexterous Manipulation from a Single Ego-Centric Video with Contact-Aware Optimization EgoCPT on the single-video dexterous-retargeting axis but keeping force first-class rather than optimizing kinematics against an object-motion target. The tactile-tracker ablation (hammer-lift 1/10→8/10 position-only vs with-tracker, per author thread) is the sharpest filed quantitative evidence that force set-points without a tracker are not sufficient — the controller side is load-bearing.
• [2026-10-28] RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions: RynnVLA-Latent (Alibaba DAMO; Zhao, Zhao, Su et al.) stakes a 3D-grounded continuous latent-action channel position on the interface-location axis: RynnLAM combines residual future-feature prediction with camera-compensated 3D-flow supervision so human + robot + sim video contribute compatible gradients to one shared latent space, then a 4B flow-matching policy predicts latent-action chunks before light embodiment-specific adaptation. Sibling to World Action Learning via Interaction-Centric Spectral Latent Guidance WING on interaction-centric debiasing (both subtract camera motion before targeting), but WING restricts the shared target to low-K DCT coefficients while RynnVLA-Latent keeps the full continuous latent in 3D. Trained on 30.1K hours → 745.52M latent-action descriptors; reports 98.8% LIBERO / 88.6% RoboTwin 2.0 / 51.9% VLABench average, and highest observed success among compared methods on six real-world dexterous-hand + mobile-gripper tasks. Middle operating point between EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale’s 22-DoF joint retargeting at ~21K h and Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2’s wrist-only reduction at 1M+ h — same corpus scale as EgoScale but with a learned embodiment-agnostic interface rather than either joint-space retargeting or hand-engineered DoF reduction.
• [2026-10-28] Hand-in-the-Loop: Improving VLA Policies for Dexterous Manipulation via Seamless Hand-Arm Intervention: HandITL (Li, Huang, Xu et al.; SJTU + ByteDance Seed + collaborators) stakes out a relative-retargeting-anchored-at-intervention-timestamp position on the interface-location axis distinct from physics-in-outer-loop (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting ReActor), kinematic 4D reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos Do as I Do), object-tracking-as-constraint (ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control ConTrack), reduced-DoF bridging-action (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots Translation-as-Bridging-Action), and hardware-locked kinematic equivalence (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic): keep the full dexterous channel but drop the absolute-pose-match assumption at the takeover moment, instead optimizing per-fingertip incremental motion transfer under pinch-reinforcement/kinematic-safety/velocity constraints, with operator wrist motion routed as EMA-smoothed residual twists on top of policy-predicted arm commands. Scoped specifically to the online-intervention / IIL regime (as opposed to offline-replay substrates like WARP, or offline mass-video substrates like Natura-1 / EgoScale / Dyna-2) and to bimanual high-DoF dexterous hands where gesture jumps become the limiting correction-pipeline artifact. Reports takeover jitter −99.8%, grasp-failures-post-takeover −87.5%, completion-time −19.1%, and discontinuity reduced by two orders of magnitude vs direct teleoperation takeover; downstream policies trained on HandITL correction data beat teleop-trained baselines by 19% average on three long-horizon dexterous tasks. First filed instance to isolate takeover-moment retargeting as the retargeting-quality lever specifically for online human-in-the-loop VLA refinement rather than for offline motion-corpus construction.
• [2026-10-27] AHa-3D: Agentic Tool Use for Real2Sim with GPT-6 Astra: AHa-3D reconstructs human motion from a casual indoor video, places it in a shared Blender scene with foot-ground and object contact refinement, and supports generating new actions for people in the scene — adjacent to retargeting pipelines in that it provides contact-refined human motion in a simulation-ready scene, but with video re-rendering rather than robot control as the downstream payload.
• [2026-10-26] Agentic World Action Models: AW-0 handles the human→robot interface at the program-vocabulary granularity rather than the action, motion, or latent-intent levels — the annotation pipeline distills a separate 27-skill library for egocentric human data (vs 23 for YAM bimanual robot data) from 1,000 h of egocentric human manipulation, with Gemini 3.1 Pro authoring per-episode programs and Claude Opus 4.8 consolidating recurring routines. Both libraries are frozen and the corpora are jointly fine-tuned under the same code+video+action denoising objective, with the memory cache absorbing episode-level plan history. Distinct from the retargeted-action lineage (Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data Ego2Robot, HuRo: Robotizing Human Videos for Scalable VLA Pretraining HuRo) and the shared-latent-intent lineage (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID, Learning What to Say to Your VLA: Mostly Harmless Vision Language Action Model Steering LFP) by sharing program syntax across embodiments while permitting embodiment-specific skill vocabularies — a middle position between fully-shared and fully-embodiment-specific interfaces. Separate skill counts (23 vs 27) is itself a datapoint on the measured task-primitive overlap between a parallel-gripper bimanual arm setup and bare-hand egocentric manipulation at matched annotation methodology.
• [2026-10-23] DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations: DONIC (Sun, Cao, Fei et al.; NTU + collaborators) stakes a bimanual-specific position on the interface-location axis: keep each hand’s reference-dictated contact role as the per-hand supervision signal while object-tracking reward stays active throughout, so one hand’s progress never masks the other’s drift. Trained jointly across heterogeneous human-motion sources (BiHOI’s 32-camera multi-view captures of 1,575 sequences × 229 objects, retargeted to robot hands, plus public mocap), a single checkpoint reaches 86.79% success on unseen trajectories of training objects and 57.32% on held-out objects without fine-tuning, surpassing bimanual-adapted DexTrack and Dexplore baselines by 77.36 and 45.73 points under a stricter two-hand metric (worse-hand error, both-hands-meet-thresholds success). Direct bimanual extension of ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control ConTrack’s single-hand object-tracking-as-primary-signal recipe, and sibling to Do as I Do: Dexterous Manipulation Data from Everyday Human Videos DO AS I DO on the “bypass teleop, convert human video” axis — but with the retargeting target shifted from a single hand to coordinated bimanual contact roles. The 77-point margin over bimanual-adapted single-hand baselines is the sharpest filed evidence that bimanual coordination needs dedicated reward structure (per-hand contact-role supervision + clip-aware advantage normalization) rather than scaling single-hand recipes by doubling the action space. BiHOI dataset (1,575 sequences, 229 objects, 32-camera 1920×1232 @ 30 fps) is a new mocap-grade bimanual HOI substrate at the capture-fidelity end that single-view reconstruction pipelines (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos, EgoCPT: Learning Dexterous Manipulation from a Single Ego-Centric Video with Contact-Aware Optimization) cannot reach.
• [2026-10-21] SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model: SuperMimic (Berkeley; Liao, Shao, Long, Zakka, Horowitz, Abbeel, Sreenath) sidesteps the interface-location debate at training time and instead changes the economic lever: a differentiable world model + BPTT motion-tracking policy co-trained on logged rigid-body-simulator transitions reports >1000× GPU-hour reduction vs a SOTA motion-tracker baseline (3 h / 1 GPU surpasses 7 days / 128 GPUs; 25 failures / 129,514 motions vs 179; zero failures at 9.8 h / 2 GPUs), with world-model prediction error flat from 1 to 264 hours of motion. The retargeting pipeline that produces the 129,514 reference motions lives upstream and is unchanged — but the compute savings mean the retargeted-corpus axis can grow orders of magnitude without the policy-training cost scaling with it, a different lever on the retargeting-quality-vs-retargeting-diversity trade-off than Compute-Optimal Scaling Laws for Human Motion Generation Cartwheel’s data-side scaling law, Scaling Behavior Foundation Model for Humanoid Robots ScaleBFM’s PPO + source-diversity recipe (model-free; 128-GPU regime), or WARP: Whole-body Retargeting from Offline Human Motion via Closed-Form c-SEW WARP’s closed-form c-SEW IK upstream fix. Real-world Unitree G1 deployment across diverse acrobatic motions.
• [2026-10-21] Learning from Fleet Experience: Collective Experience Learning across 96 Robots: CEL/FERT (Tsinghua + AgiBot + Infinigence AI) is a counterpoint to the human-video / egocentric-pretraining lines on this page: after an SFT + (M)HG-DAgger initialization that still consumes human demonstrations and expert interventions, the self-improvement loop is driven entirely by robot-generated autonomous rollouts, with humans supplying only binary success/failure labels and physical environment resets. Reaches 97.3% mean success across five real-world bimanual AgiBot G2 tasks (and 93.3–99.2% on two robots excluded from training data collection) at 96-robot fleet scale. Useful datapoint on the operational trade-off between scaling the human-side data substrate (teleop corpora, UMI, egocentric human video, wearables) and scaling the robot-side fleet that generates its own training distribution from autonomous rollouts — on this page’s taxonomy, FERT argues that at tabletop bimanual industrial scope + sufficient fleet size + a sufficiently good SFT+DAgger initialization, the per-task self-improvement loop does not need any additional human-side retargeting substrate to climb from 81.0% to 97.3% mean success.
• [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) stakes out a shared-channel = future-force + 2D visual tracks position on the interface-location axis: use 2D visual tracks (plus future video) as the embodiment-agnostic task representation to bridge human and robot sides, and bind force prediction into coupled coarse/refiner training on top. Pretrains on EgoForce (large-scale human manipulation corpus with recorded contact forces, released with the paper) — the first filed foundation-scale human corpus where force is a first-class channel rather than inferred from vision (Being-H0.8: A Latent Tactile World-Action Model at Scale Being-H0.8) or captured via tactile gloves (TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning TouchScale, ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine ACE-Data-0). Sibling to T-Rex (T-Rex: Tactile-Reactive Dexterous Manipulation) as the headline baseline, beaten by up to 60 pp on ten contact-rich force-sensitive tasks (tool use, musical instruments, precision lab procedures) with only 10 robot + 90 human trajectories at fine-tuning. Load-bearing ablation evidence: force supervision in the human pretraining stream carries the gain — removing it collapses the downstream advantage. Human:robot ratio sweep argues the human-video channel transfers nearly as effectively as robot demos under this recipe, a sharper quantitative statement than the “human data helps” framing of HuRo / Ego2Robot. Complementary 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 VLA force modulation) on the integration-pattern axis.
• [2026-10-19] Has GPT-6-Astra Achieved Embodied Intelligence?: SenseNova Joint Research Team measures the no-explicit-retargeting-step pattern (Astra consuming ICL demonstrations directly) at breadth-axis scale: 12 benchmarks × 8 real robot configurations expose that per-embodiment ICL-demo collection cost becomes the dominant blocker once standalone Astra numbers are translated to physical deployment.
• [2026-10-18] EgoCPT: Learning Dexterous Manipulation from a Single Ego-Centric Video with Contact-Aware Optimization: EgoCPT (Sun, Shi, Ren, Yang, Cutkosky, Xu — Stanford + collaborators) stakes a per-segment contact-aware optimization position on the interface-location axis: fix object motion as the retargeting target and treat the demonstrated hand motion as a search prior, then run an inner loop (sample reachable contacts → refine pose to remove penetration/gaps → allocate dynamically-feasible contact forces for the target 6-DoF object motion). Single-video RGB + reference hand-pose input; reports 67.3% average success on EgoDex (noisy monocular hand-pose) and 92.0% on TACO (mocap-grade) across five manipulation tasks (pouring, lint rolling, sweeping, scooping, pick & place; plus bowl & ladle, container manipulation, surface brushing). Closest sibling to C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video C2Dex (object-side contacts as shared interface) but differs by attacking the problem per-interaction-segment via geometric+dynamic contact-force optimization rather than cross-frame aggregation + Laplacian preservation + residual RL; sibling to CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation CHORD at the planning-time rather than policy-reward layer, and complement to Do as I Do: Dexterous Manipulation Data from Everyday Human Videos Do as I Do by moving the retargeting target one layer out from hand kinematics to object dynamics. The ~25-point noisy-vs-clean gap (EgoDex vs TACO) quantifies residual sensitivity to upstream hand-pose quality the optimization can’t fully absorb.
• [2026-10-15] HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution: ToolBook (companion to HumanoidToolBench) releases 3,003 Meta-Quest-teleoperated Unitree G1 simulation demos + 91 real-robot teleop demos across 55 tool meshes under CC BY-NC 4.0 — eight teleoperators map onto upper body + walking primitives, with every rollout labeled against a five-event sub-success chain so retargeting policies can be supervised at sub-task granularity rather than only on final success. Decoy mode (same-category confusing tools) makes the corpus a candidate substrate for functional-affordance disambiguation under retargeting.
• [2026-10-13] InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data: InternW0-Δ integrates Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data Ego2Robot rendered-from-human-video data as one of four streams in its unified >20K-h open corpus (alongside robot demos, UMI, and raw egocentric human) under a canonical state-action representation with a shared semantic layout (joint positions, absolute EEF pose, gripper/hand configuration), with EEF actions relative-to-current and other channels absolute. Takes the “absorb the gap in the training objective” position by jointly training on all four streams under the same WAM objective rather than separating human-video pretraining from robot post-training — distinct from ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining ACE-Ego-0’s reliability-aware joint training (which models per-sample pseudo-label noise explicitly) and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale’s scale-the-egocentric-pretraining answer by putting the retargeting weight on the canonical representation and source mixture layer rather than the loss or the collection rig.
• [2026-10-13] The Will to Act — Natura-1: Natura-1 (Nature Will) joins the “pay for retargeting noise with scale” branch at the 1M-hour egocentric-human-video mark — directly comparable to Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2 at matched scale but with the full 20-DoF anthropomorphic-hand action channel preserved rather than Dyna-2’s wrist-only reduction, and ~48× EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale on the same high-DoF substrate. Stakes the position that anthropomorphic hands are the “closest correspondence available today” to the human hand pretraining data is grounded in, so the retargeting interface stays at full 20-DoF joint configuration rather than reducing dimensionality (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), swapping to object-side contact topology (C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video), or routing through task intent (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition). Load-bearing infrastructure claim: in-house motion-reconstruction pipeline achieves 41% MPJPE reduction vs existing open-source approaches on an internal benchmark with mocap ground truth, which is the retargeting-quality counterpart to the retargeting-scale lever — argues the 1M-hour corpus is only usable because the per-frame hand/body pose labels are mocap-grade rather than noisy. Scaling evidence (44K → 1M hours → 67.7% MSE drop on zero-shot real-robot action prediction) is the first quantitative support filed for human-video-to-dexterous-hand transfer at this scale.
• [2026-10-13] Jiahua Dong — 30,000-hour egocentric world model learns hands cleanly, not objects; neither 100× data nor 4× model closes the gap: Jiahua Dong trains a world model on 30,000 h of first-person Mecka data and reports pre-training beats released Cosmos 3 across all metrics on downstream humanoid transfer — first filed public datapoint pitting a scaled human-video pretraining substrate against a released frontier WFM baseline rather than a VLA or internal control. Adds a channel-choice ablation on the input side: giving the model 3D hand pose as input saturates hand prediction with a fraction of the raw-video data — echoing EgoScale’s finding that what is retargeted matters, not just hour count, now on the pretraining-input side rather than the action-channel side.
• [2026-10-12] TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning: TouchScale (Li et al.) adds a tactile-channel-at-scale position to the interface-location debate on this page: shared human→robot interface is a 30 Hz time-aligned vision + wrist-vision + 880-taxel-per-hand bimanual tactile tuple, with a future-tactile-change prediction mid-training objective bridging the human corpus to the downstream xArm6 + BrainCo Revo 2 policy. 500 h / ~2K task descriptions on a single unified wearable rig; MIT-licensed data + code. Reports real-world success lifting 22.5% → 57.5% average on four contact-rich tasks (bottle-cap removal, test-tube transfer, whiteboard wipe, soft/hard sorting) under matched 50-demo / 20-trial-per-task evaluation. Sibling to ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine ACE-Data-0’s tactile-glove + mocap recipe but single-rig / consistent-protocol at larger scale, and complementary to Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1’s UMI-based scale bet by moving the consistent-sensing argument from gripper geometry to tactile channel.
• [2026-10-12] Yeah Hand — a mostly-3D-printed anthropomorphic hand with tendon drive, ROS, and a prosthetic variant: Yeah Hand (Vittorio Lumare, Tindie beta) drops the open-anthropomorphic-hand BOM to ~$217 kit / ~$581 assembled with a 5-servo fishing-line-tendon design, forearm-mounted motors, flex-sensor teleop glove, and a prosthetic variant — sitting at the hobbyist-cheap / low-DoF extreme of the open-hand-hardware axis below ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning (17-DoF) and Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand (16 joints / 7 motors, $314). Vendor-stated 800-cycle reliability test is ~1000× below ORCA / Aero Hand lab benchmarks, so “research-grade” is relative to hobbyist baselines; the structural contribution is lowering the activation energy for small teams and demonstrating the robotics + prosthetic product-line variant split.
• [2026-10-10] Neocambrian AI launches Mitts — tactile gloves for high-precision robot touch data: Neocambrian AI launches Mitts — wearable tactile gloves (450+ sensing points, 5 g sensitivity, 30 Hz) synced with phone/RoboCap video, pitched as human-demonstration capture for robot manipulation. Collection-hardware-first sibling to glove-based substrates (ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine ACE-Data-0) and bimanual force-paired corpora (Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge Sharpa Origami), adding the whole-hand tactile channel to human-video capture rigs that have been overwhelmingly RGB + hand-pose until now. Announcement only — no disclosure of which target robot embodiment the captured data is retargeted onto or how the human-tactile stream maps to robot-hand sensors.
• [2026-10-10] A Balanced Data Diet: Addressing the Exploration Bottleneck in Mega-Scale RL for Robot Control: SGS stakes out the “no human demonstrations at all” extreme on the retargeting axis — pure sim PPO + adaptive task sampler produces dexterous assembly policies that zero-shot to real UR5e via RGB distillation, with emergent strategies suited to the robot’s own gripper rather than anthropomorphic imitations (tipping a rod against the board leg, uprighting a gear against the shaft, flipping a nut in place instead of regrasping). Counter-position to the retargeting-centric dexterous-manipulation lineage (CHORD, DemoMimic, ConTrack, “Do as I Do”) this concept page tracks — argues that when the simulator is high-fidelity enough and the sampler finds the frontier, the kinematic/morphology gap retargeting tries to bridge can be sidestepped entirely by letting RL find gripper-native strategies from scratch.
• [2026-10-09] In-context Robot Learning Made Simple: A Democratized Recipe for Manipulation Tasks: SimpleICL (Alibaba Qwen + CASIA) stakes a problem-definition position: no explicit retargeting step, keep the raw human demo video in-context, bottleneck through fixed-length query tokens that cross-attend to a frozen ViT over the demo and are injected into both video and action DiT cross-attention. The paper’s actual contribution is a four-dimension decomposition of what the robot should follow (action / composition / object / affordance) + three robustness axes it should ignore (spatial mismatch / human-detail / scene variation), which the collection-group data protocol then instantiates via within-scene discrimination pairs. Sibling to Behavior Prompting Policy: Demonstrations as Prompts for Manipulation BPP on the attention-pooled-chunks architecture axis but WAM-mounted rather than sensorimotor, and to Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization Zero-WAM / S1: In-Context Learning for Robotics S1 / GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5 on the no-explicit-retargeting-step branch — distinct in attacking the problem via problem-definition + low-cost data recipe (3,000 pairs/robot/day) rather than scale, synthetic data, or a dedicated training objective.
• [2026-10-09] UniWAM: Unified World-Action Model: UniWAM (Song, Chen, Wang et al.) reports the first filed log-linear scaling law for unified human-robot co-training — distinct from EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale’s human-video-only scaling, Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2’s million-hour human-video + cross-embodiment transfer, and Isaac 0.5: An Open-Weight Embodied Foundation Model Isaac 0.5’s measured video-vs-teleop exchange at a fixed 80:30:30 composition — by characterizing scaling on the mixture itself under a joint-training objective (physical reasoner + world generator + action predictor on one backbone) rather than a staged pretrain-then-finetune recipe. Trained on 10,000+ hours of robot + human-egocentric + VQA; a complementary-supervision pretraining recipe routes each stream to the heads it benefits most (VQA → reasoner, egocentric → reasoner + world generator, robot → action predictor + world generator). Low-level actions tokenized as natural language to preserve VL capability. Apache 2.0 open code + ModelScope weights. Sharpens the open question on the page — is the mixture axis a distinct scaling lever — with a quantitative “yes, log-linear in co-training ratio” datapoint.
• [2026-10-08] RoboJEPA: Scaling Robotic Latent World Models: RoboJEPA pretrains a single JEPA predictor across 12 robot embodiments of real-robot data and deploys zero-shot on real hardware via goal-image planning, with reported predictable compute-scaling of both imagination error and downstream planning performance. Multi-embodiment sibling to the cross-embodiment scaling story in Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models Dyna-2 (human-video pretraining transferring to YAM robots) and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale (humans-as-another-embodiment framing), but on the JEPA latent-predictive side rather than the generative-rollout side.
• [2026-10-08] EigenDEXplore: Structured Exploration for Dexterous Manipulation with Human Priors: EigenDEXplore (Gupta, Lum, Wang, Pan, Liu, Bohg, Song — Stanford) stakes a new position on the interface-location axis: the human prior enters the RL exploration-noise distribution rather than the action space, reward, or reference trajectory. PCA on retargeted human-hand trajectories yields a low-DoF eigenbasis that shapes noise added to policy actions, while the mean action retains full robot joint DoF. Drops into three different pipelines (SimToolReal object-centric tool use on Sharpa+KUKA, DeXtreme Allegro in-hand reorientation, DexMachina reference-guided bimanual RL with task reward only) and SPIDER cold-start trajectory optimization. The lift is largest on highest-DoF hands under both DexMachina and SPIDER. Sibling to REGRIND: A Minimalist Retargeting-Guided RL Recipe for Dexterous Manipulation REGRIND (prior as reference) and complementary to ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting ReActor as a composable exploration primitive — the mean-action-vs-noise separation is the load-bearing design choice distinguishing it from action-space-reduction recipes like Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots.
• [2026-10-08] Long-WAM: Scaling the Context of World-Action Models: Long-WAM’s robot-domain AR continued-pretraining corpus blends robot video (RoVid-X, AgiBot World) with egocentric human video (EgoDex, EgoVerse, VITRA) at ~10k window-equivalent hours, teacher-forced next-chunk prediction up to 30s with no action labels — and reports this mix lifts peak success on both RoboCasa GR-1 and LIBERO-Long. Another vote for the ego-video-as-pretraining-fuel thesis (cf. ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining, EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data, HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining) with the additional structural requirement that the pretraining objective be causal/autoregressive.
• [2026-10-07] World Action Learning via Interaction-Centric Spectral Latent Guidance: WING (Liu, Miao, Chen et al.; HKUST + collaborators) stakes out a new frequency-domain position on the interface-location axis: the shared human-robot interface is the lowest-K DCT coefficients of an interaction-centric latent-action sequence, used as a soft guidance target the robot action generator attends to via gated residual updates — not a reduced action space (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), not 4D kinematic reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), not a shared joint channel at scale (EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data). Two load-bearing design choices: (a) interaction-centric debiasing via point tracks + hand-object masks removes ego-camera motion from the latent-action encoder before any downstream use, directly countering the “pay for noisy latent actions with scale” recipe; (b) DCT truncation (K=4 Pareto-best, K=2 under-dynamics, K=8 noise-reintroducing) empirically localizes the cross-embodiment signal to slowly varying temporal structure. Reports 99.20% LIBERO / 93.80% RoboTwin 2.0 / 57.7% RoboCasa-GR1 average success + strong real-world generalization, with ablations showing predicted spectral guidance beats both video-only pretraining and direct action-target pretraining at matched 200 h egocentric budget. Sibling to LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID (task-intent soft guidance) and WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time WAM-TTT (KV-memory interface) on the soft-interface axis but with the retargeting signal specified in a transform domain rather than a 3D-scene or opaque-cache domain.
• [2026-10-01] Argus: An Open-Source Annotator for Robotics Data: Argus audits FastUMI-100K and HABIT (both on the UMI/human-ego substrate this page tracks) with VLM-based semantic checks beyond Pantheon’s prior deterministic sweep: 1.5% of FastUMI episodes have swapped left/right camera streams, 6% have recorded motion disagreeing with video, 1.7% have recorded gripper openings that never happen; HABIT has 3.8% privacy-blurred donuts (task objects obscured by face-blur post-processing). Deterministic motion-correlation alone cannot find all swaps because 64 FastUMI episodes have broken motion data, so Argus uses the VLM to detect swaps directly from footage — a datapoint for how far the “trust the recorded pose” assumption breaks on UMI capture rigs.
• [2026-09-30] DemoMimic: One Demonstration, Many Objects — Generalizing Manipulation via Local Contact Geometry: DemoMimic (Stanford; Sharma, Sahoo, Huang, Fei-Fei Li, Wu, Sadigh, Bohg) stakes a new position on the interface-location axis: the shared interface is local contact geometry at the interaction site, and retargeting error is absorbed at the reward layer inside a single-video sim RL run — an Alignment Reward (surface-normal alignment between hand link and object at contact) plus a Sustained Contact Reward (quadratic bonus on contact streaks) keep contacts stable and continuous, closing the sim-to-real gap that reward ablations and baselines (DexMachina*, HERMES*) fail on. Reports 71% average real-world success across 16 objects × 4 tasks × 2 hand embodiments with the smallest sim-to-real gap among baselines. Complement (not alternative) to same-authorship DexAgent: An Agentic Human2Sim2Robot Framework for Dexterous Manipulation with Self-Evolving Tool Library DexAgent — DexAgent fixes the H2R pipeline at the tool-library layer, DemoMimic fixes it at the reward layer inside RL; orthogonal levers on the same underlying problem. Sibling to CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation CHORD on the contact-interface axis (geometric analog to CHORD’s wrench-based interface) and to Do as I Do: Dexterous Manipulation Data from Everyday Human Videos DO AS I DO on the single-monocular-source axis (RL-in-sim rather than kinematic reconstruction).
• [2026-09-29] DexAgent: An Agentic Human2Sim2Robot Framework for Dexterous Manipulation with Self-Evolving Tool Library: DexAgent (Wang, Li, Fei-Fei, Wu, Huang; Stanford) proposes an agent-authored Human2Sim2Robot pipeline: an LLM agent runs the four stages (video understanding → property-based sim reconstruction → trajectory optimization → data generation) but selects tools from a persistent library per object property (rigid/articulated/deformable) and writes new tools when the library falls short, with property-specific verifiers gating each stage before propagation. Structurally new position on this concept: prior filings (DO AS I DO, Ego2Robot, HuRo, C2Dex) run fixed pipelines and vary the interface; DexAgent varies the pipeline itself per video via agent tool selection, and reports 3.5× success rate over baselines across 11 real-world tasks. Closest sibling to Introducing Auto-Engineering for Robotics GRID’s Video-to-Sim sub-mode, but scoped tightly to the H2R2 data-generation subproblem with explicit verifier-gate discipline.
• [2026-09-28] Chestnut Robotics launches Aero UMI + Aero Hand — a co-designed capture rig + high-DOF dexterous hand system: Chestnut Robotics launches Aero UMI + Aero Hand at IROS 2026 as a co-designed wearable capture rig + high-DOF dexterous hand, arguing that prior UMI variants scaled data collection without ever co-designing the collection tool with a dexterous hand at matching DOF — so recorded data never fully mapped to hand capabilities. Sibling on the hardware-first co-design branch to mimic (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1), 15-DoF M1+U1), TwinDEX (TwinDEX: A Scalable Path from Robot-Free Data Collection to Real-World Dexterous Manipulation, bimanual 9-DoF), and RAI Koala (Getting a Grip on Robotic Data Collection — Co-Designed Koala Gripper for Force-Mediated Manipulation, parallel-jaw), distinguished by explicitly pushing the co-design constraint into dexterous rather than parallel-gripper endpoints. Launch surface only — no DOF, BOM, or empirical numbers disclosed.
• [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: Δ₀ bakes retargeting into pretraining rather than adding it downstream — 10,000+ h of paired egocentric observations + whole-body motion pretrained in a shared action space across embodiments, via the Delta D1 head-mounted panoramic capture device. Direct alternative to per-embodiment specialization or downstream retargeting stages: the invariance lives in the pretraining action-space design.
• [2026-09-28] We Looked at the Data — Diagnosing Open-Source Robotics Datasets for WAM Training: Pantheon audits the UMI substrate the “make-collection-tool-embodiment-invariant” recipe branch depends on — FastUMI_100k_lerobot, HiFi-UMI-2K, XDOF/ABC-130k — and documents anomaly rates that vary structurally across datasets (missing frames concentrate in ABC-130k, mutual-gripper pixelation is a HiFi-UMI-2K artifact, spiky per-frame actions are dataset-uniform in UMI corpora specifically because they’re not constrained by robot arm kinematics). Sharpens the HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone HiFi-UMI fidelity-first framing and the Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1 scale-first framing by supplying a shared audit substrate and SE(3)-correct action-smoothing pipeline downstream of both.
• [2026-09-28] EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics: EmbodiedSWE ships explicit cross-embodiment support (Franka, xArm7, Kinova Gen3 + panda hand, Unitree G1, bimanual “multi”) behind a common BaseRobot + controller-zoo interface, with the paper claiming that coding-agent solutions transfer across both tasks and embodiments. Positions coding-agent-authored trajectories as an alternative to human-video / mocap retargeting for producing multi-embodiment training data — the “kinematic gap” is closed by re-solving the task with the agent given a new embodiment adapter rather than by projecting recorded human motion onto a target morphology.
• [2026-09-28] HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own: HomeBody sidesteps human-to-robot retargeting entirely by using the robot’s own exploration data (ego video + SLAM + joint pose + waypoints) as the substrate for spatial reasoning — the same GPT Astra that acts as the high-level policy also authors an Isaac Sim digital twin of the unseen kitchen from that data, and every downstream skill call resolves through analytic IK + minimum-jerk planning + collision-check on the humanoid’s own kinematics. Complement (and contrast) to the human-video-as-primary-substrate recipes (HuRo: Robotizing Human Videos for Scalable VLA Pretraining HuRo, ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining ACE-Ego-0, Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness Astra+human-video) — HomeBody shows that when the LLM is powerful enough to author its own spatial memory, the retargeting problem can be routed around by staying entirely in the robot’s own embodiment throughout.
• [2026-09-23] Data Scaling Laws in Imitation Learning for Robotic Manipulation: Lin, Hu, Sheng, Wen, You, Gao (Tsinghua + Shanghai Qi Zhi, Oct 2024) is the foundational scaling study on the “make-collection-tool-embodiment-invariant” recipe branch of this page — UMI hand-held grippers are the retargeting-at-collection-time device, and the paper establishes that the effective scaling axis is diversity (environments × objects) rather than raw demonstration count. Sibling in substrate to Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1 (100K UMI hours), Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic (M1 hand + U1 wearable), and HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet) HandUMI, but operates at the small-team / single-task-generalist regime with a specific quantitative recipe: 32 environment-object pairs × 50 demos, one object per environment. Adds the earliest quantitative anchor to the “collection-tool-invariance is a valid retargeting strategy at small scale” position that later UMI-scale bets extend.
• [2026-09-21] LeRobot — Eidon AI winds down and open-sources 9TB egocentric housework dataset (tracker-pov) with 7-point IMU arm tracking: Eidon AI’s shutdown data drop pairs first-person housework video with 7-point IMU arm tracking — a distinctive sensor combination (not full-body pose, not hand pose) directly relevant to arm-only retargeting pipelines from human wearables to robot embodiments.
• [2026-09-20] CausalWM: Causal Chain-of-Thought Reasoning for Embodied World Model: CausalWM’s training substrate blends ~31K source hours (~20K retained after curation) of human egocentric video (Ego4D, Egocentric-10K, EgoDex, EPIC-Kitchens, H2O, EgoVerse), real-robot data (AgiBot, Galaxea, RoboCOIN, RoboMIND, DROID, RT-1, BridgeData V2, OXE, Humanoid-Everyday, RoVid-X), and simulation (InternData-A1, RoboCasa365, RoboTwin 2.0, AgiBot-World 2026 digital twins). Same substrate-mixing pattern as HumanScale / EgoScale / Dyna-2, with a source-adaptive filtering pipeline (motion thresholds, actor/interaction verification, contiguous event assembly) rather than uniform preprocessing.
• [2026-09-18] PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics: PointZero (Duisterhof, Zhang, Hung, Wen, Birchfield, Li, Ramanan, Ichnowski — CMU + NVIDIA + Columbia) adds a 3D-scene-dynamics variant of the action-free-pretrain-then-add-action-head recipe: pretrain on dense 3D point track completion over 2.9M synthetic frames (no robot actions anywhere in the loss), then post-train with 20 action-labeled demos + 100 actionless demonstration videos per task. Beats DP3 on all 7 sim + real manipulation tasks and matches or exceeds 3PoinTr on 6/7 (Blockstack 99.8, Microwave 93.1, Drawer real 100, Cup real 100). Sibling to LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID’s task-intent retargeting (short-horizon 3D object flow + palm-pose reference) but scaled up to full dense scene dynamics as the pretraining substrate, and to μ₀: A Scalable 3D Interaction-Trace World Model µ₀ (frozen trace-WM + action expert) with tracks moved from downstream interface into pretraining supervision. The load-bearing evidence for retargeting: pretrained → PointZero-FT beats PointZero-Scratch by 2–3× on 4/6 real-robot PGND dynamics scenes and by +7.7 pp avg (up to +12.9 pp on Glass) in IL success at matched 20-demo budget — pretraining is where the transferability lives.
• [2026-09-18] ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning: ORCA (Christoph et al., ETH Zurich Soft Robotics Lab; IROS 2025) is the target-embodiment substrate side of the retargeting pipeline: 17-DoF anthropomorphic tendon-driven hand for <2,000 CHF and <8h assembly, with mocap-glove demonstrations retargeted via energy-based minimization onto the hand’s joint space to drive a diffusion-transformer policy that runs 7h 17min of autonomous pick-and-place. Sibling to Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand Aero Hand Open on the open-hardware-hand axis but at higher DoF and dedicated fingertip tactile; complementary to Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots Translation-as-Bridging-Action’s reduce-the-action-space recipe by providing an anthropomorphic target that keeps the full articulation channel available.
• [2026-09-17] Agent as Policy for Robotic Manipulation: AGP includes assembly-from-human-video as one of its evaluated configurations — a human demonstration video is dropped in as the task specification and the LLM agent drives insertion at 0.73 mm clearance via depth-verified EE-pose steps, no explicit retargeting stage. Sibling in position to Wenli Xiao — GPT-6 Astra does physical in-context learning from a single human video via a motion-planning tool harness (Astra + human-video ICL via motion-planning tool harness), pushed one step further with a controlled paper-scale evaluation showing the pattern extends to sub-millimeter precision insertion rather than only qualitative “just works” demos. Adds an LLM-in-the-loop, program-written execution row to the no-explicit-retargeting-step position already staked out by GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5, OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence DM0.5, and Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization Zero-WAM — with the retargeting happening entirely inside the LLM’s forward pass over the human video, then materialized as executable Python against the EE-pose interface.
• [2026-09-16] Modality-Autoregressive World-Action Models: ModAR reports that adding human video co-training improves its modality-autoregressive WAM’s real-world bimanual performance on top of robot data — replicating the EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data EgoWAM finding that the WAM head is the effective interface for human data to enter a policy without requiring paired human-robot action alignment. The specific modality set (point tracks + DINO + depth, no RGB) is action-morphology-agnostic by construction and the retargeting channel is implicit in the WAM training rather than requiring explicit action-space alignment.
• [2026-09-14] RewardAI OM-1 + Omnibody Hand launch — zero-shot cross-embodiment dexterous manipulation from human data (DexCap successor): RewardAI OM-1 launch positions itself as the successor to DexCap (chest-mounted RGB-D + SLAM + EMF-glove mocap → LEAP-hand retargeting via IK + Diffusion Policy on retargeted point clouds), retargeting instead onto a new 7-DoF Omnibody Hand rather than a stock LEAP hand. Sibling on the glove-based human-mocap → dexterous hand axis to Do as I Do: Dexterous Manipulation Data from Everyday Human Videos Do as I Do (kinematic 4D-reconstruction pipeline) and HUG: Human Universal Grasping HUG (MANO articulation preserved, capture-quality-focused) but distinct in shipping the target-embodiment hand hardware alongside the policy. Announcement surface only — no retargeting-recipe details, no quantitative comparison to prior filed positions (physics-in-loop / kinematic-reconstruction / RL-constraint / DoF-reduction / hardware-lock).
• [2026-09-14] HuRo: Robotizing Human Videos for Scalable VLA Pretraining: HuRo (Yonsei; Jeong et al.) sits on the “robotize both observations + actions and pay for retargeting-noise with scale” branch alongside Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data Ego2Robot: a single pipeline that visually transforms human videos into robot-style observations and retargets hand motion into robot actions, inferring missing intermediate signals across annotation levels so five heterogeneous sources aggregate into 630K episodes / 142M frames. Distinct from Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s drop-the-noisy-channels recipe by keeping the full action channel and from EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale’s single-source axis by cross-source aggregation. Load-bearing ablation: end-to-end pretraining with action retargeting beats visual-only transfer, and visual robotization specifically improves OOD robustness — the +37.3 pp OOD gain (34.9→72.2%) exceeds the +28.8 pp in-distribution gain (51.5→80.3%) across four real-world tasks. No per-source or per-task numbers in the abstract.
• [2026-09-10] Introducing Auto-Engineering for Robotics: GRID’s Video-to-Sim sub-mode collapses the human-video-to-robot pipeline into a single skill-creation option the agent picks when robot demos aren’t allowed: hand-pose estimation + 3D object segmentation from one phone video → 3D reconstruction of hand + object motion → interaction reconstructed in simulation → varied and checked for physics consistency → synthetic BC demos. The physics simulator plays the role of the retargeting engine (rather than a learned intent module or dedicated retargeting model as in HumanScale / EgoInfinity / ConTrack / DO AS I DO / LUCID / HOST), and the illustrated task (swirling liquid in a flask) is chosen specifically because conventional teleop can’t reproduce the required rapid coordinated circular motion.
• [2026-09-10] SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators: SyncWorld frames cross-embodiment / cross-viewpoint transfer as an Action–Visual Mapping inference problem solved by a short visual calibration episode in context, rather than by explicit kinematic retargeting — the same pretrained DiT then rolls out on xArm, ARX5, KUKA, Franka Panda, and UR5e without weight updates, including embodiments absent from training.
• [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 pushes the no-explicit-retargeting-step position (already held by GEN-1.5 in-scene physical prompt, DM0.5 in-context human demo, Zero-WAM in-context prompt) one abstraction level higher: no VLA at all, and the human demonstration is a pre-recorded video dropped into the codex app rather than a live demo in the robot’s scene. GPT-6 Astra emits EE-pose tool calls, a motion-planning + IK harness executes them — the entire human-motion → robot-motion retargeting collapses into the LLM’s in-context understanding, with no designed representation (no 3D flow, no task-progress manifold, no reduced action space). Qualitative first-pass claim only, no per-task breakdown.
• [2026-09-09] CosmoH2G: A Hand-to-Gripper Transfer Dataset and Baseline Method for Object Manipulation with Complex Spatial Movements: CosmoH2G (Zhao et al., CUHKSZ + FNii-Shenzhen + GenuX; SIGGRAPH Asia 2026) stakes out a new position on the interface-location axis specifically targeting complex spatial trajectories (rotations, flips) where existing kinematic retargeting collapses from compounding drift: a two-stage decomposition (Stage I predicts sparse initial+terminal gripper keyframes; Stage II fills the continuous action sequence conditioned on them) with an asymmetric orientation-vs-translation split — keep orientation end-to-end learned but post-optimize translation via grasping heuristics + kinematic consistency. Sibling to Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots on the action-representation axis but distinct in keeping the full channel and drift-correcting rather than dropping the noisy dimensions; on the collection-tool side sits in the same handheld-gripper lineage as mimic (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1)) and XR-1 (Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training) but scoped to motion-complexity-curated 6,189 episodes / 1,254 objects with an explicit baseline method rather than pretraining scale. Reports “significant outperformance” of traditional baselines in sim + real-robot but no per-task numbers or ablations in the abstract; code + dataset “coming soon”.
• [2026-09-08] EgoWAM: World Action Models Beyond Pixels with In-the-Wild Egocentric Human Data: EgoWAM (Georgia Tech RL2; Baoyu Li, Danfei Xu) is the observation-side complement of the same lab’s Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots action-side recipe: after aligning actions as a strong BC baseline (14-D EE space, instantaneous device frame, embodiment-specific windows, quantile normalization), EgoWAM stakes the position that the remaining human→robot transfer happens through an auxiliary world-model channel on the shared trunk, not through a smarter action decoder. Under unaligned human data BC collapses below robot-only while the 3D-flow WAM stays robust and still surpasses robot-only — a direct datapoint that world-prediction supervision absorbs the residual action mismatch that action-level co-training amplifies. Complementary to HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining‘s matched-volume egocentric-vs-teleop head-to-head with the how rather than the whether: DINO features drive 4× OOD object/scene generalization, 3D flow drives 20–30% in-domain spatial precision, pixel-VAE targets transfer weakly.
• [2026-09-08] HDMI: Learning Interactive Humanoid Whole-Body Control from Human Videos: HDMI (Weng et al., CMU; arxiv:2509.16757, open-sourced Sep 8 2026 with mjlab backend + multi-motion support) stakes out a reward-side answer to the imperfect-retargeting problem: keep the full kinematic pipeline (GVHMR + LocoMujoco) but design a unified contact-promoting interaction reward that lets the policy deviate from flawed references to establish proper grasps, plus a residual action space that centers exploration on the reference pose rather than default standing. Sibling to ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting ReActor on the physics-in-loop axis but with the tunable knob in the RL inner loop rather than the retargeting outer loop — the two together define opposite corners of the “where to absorb kinematic-retargeting error” design space. 67 consecutive door open + traversals (34 min) and full Truman-bow sequence (climb → bow → sit → wave → jump → walk back) on Unitree G1 are the sharpest published contact-consistency numbers on this page for full whole-body loco-manipulation retargeting from monocular RGB video.
• [2026-09-06] How we actually test a robot hand: RLWRLD’s All Hands Up! archive publishes 11-level Kapandji thumb-opposition sweeps for 10+ commercial dexterous hands, measured in-simulation on each hand’s URDF — an embodiment-agnostic dexterity precondition that any human→robot retargeting recipe can check before designing the mapping (“can this hand even reach the target pose?”). Complements the recipe-level retargeting literature (physics-in-loop, kinematic-reconstruction, RL-constraint, action-representation reduction) with a target-embodiment reachability audit that operates one layer upstream.
• [2026-09-06] Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design: Transformer Transformer (Ha, Liu, Song; CoRL 2026) approaches the human→robot problem from the opposite end of every other filing on this page: instead of closing the kinematic gap between a human demonstrator and a fixed robot, take the UMI-collected end-effector trajectories as embodiment-invariant task specifications and treat the robot embodiment itself as the decision variable. The load-bearing observation is that a fixed set of end-effector trajectories “transfers unequally” to different robots — imperfect transfer becomes the co-design signal rather than a problem to eliminate. A single DiT over RoboTokens diffuses complete embodiments (links, joints, motors, inertias) + states + actions jointly under masked-modeling schemes, optimizing zero-shot for unseen rewards and unseen trajectories, and generalizes across 76 UMI trajectories (tossing, screwing, drawer-opening, scrubbing) plus UMI’s bimanual dishwashing. Fabricated ALOHA validation on Flingbot cloth unfolding cuts tracking error 73% (22→6 cm) — sibling in framing to UME (embodiment-invariant interface) and mimic/XR-1/HandUMI (embodiment-invariant collection tool) but distinct in staking the position that the follower embodiment is what should be adapted to the task.
• [2026-09-05] Universal Manipulation Exoskeleton: Learning Compliant Whole-body Policies with Real-time Torque Feedback: UME (Liang et al., Inclusion AI Ant Group + Stanford BDML) stakes out a new position on the interface-location axis: keep the demonstrator side human-shaped via a 7-DoF upper-limb exoskeleton whose 3-1-3 kinematic + coaxial actuator arrangement replicates full human wrist RoM and most shoulder RoM, and push the invariance into a sub-manipulator retargeting algorithm that decomposes both exoskeleton and follower into virtual 3-DoF spherical shoulder + 1-DoF elbow + virtual 3-DoF spherical wrist, computes per-sub-manipulator Jacobians independently, and maps kinematics + haptic torques bidirectionally using the dynamically-consistent (mass-weighted) Jacobian inverse. Independent sub-manipulator computation is the load-bearing engineering choice — it avoids the 6-D end-effector-task-space singularities that make standard task-space control brittle in constrained spaces, which is what enables the 0.95 success on 5 mm-clearance GPU picking that UMI-style end-effector-only baselines collapse to 0/20 on. Same operator drives 7-DoF OpenArm, 7-DoF Franka (sim), and 6-DoF X-ARM without per-robot retargeting redesign. Sibling in framing to Cross-Embodiment Robot Manipulation via a Unified Hand Action Space UHAS (canonical fingertip sphere for dexterous hands) but scoped to the arm level rather than the hand, and to Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots Translation-as-Bridging-Action’s DoF-reduction recipe but going the opposite direction — keep the full arm interface, push invariance into the retargeting math. Distinct from ModPack: An Extensible Teleoperation Interface for Bimanual Mobile Manipulation ModPack’s kinematically-equivalent-leader-arm-per-follower answer by using one physical exoskeleton across all follower kinematics.
• [2026-09-05] RoboTok: An Internet-Scale Data Engine for Human Demonstration Retrieval and Dexterous Manipulation Learning: RoboTok (Rice RobotPI Lab; Qian et al.) sits in the data-engine slot rather than the interface-location debate this page tracks: a motion-similarity retrieval index over web human-manipulation video, keyed on 3D hand trajectories re-expressed in an estimated actor-centered reference frame so trajectory shape describes motion relative to the person rather than the camera. Sibling to EgoInfinity: A Web-Scale 4D Hand-Object Interaction Data Engine for Any-View Robot Retargeting and Video-to-Action Robot Learning EgoInfinity’s forward-conversion pipeline but attacking the searchability axis — you retrieve motion-matched demonstrations on demand rather than curating a fixed corpus (Figure Index, HumanScale, EgoScale) — and the retrieval-side echo of Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s head-camera-frame wrist reduction: same collapse-to-embodiment-invariant-frame move, applied as a retrieval key rather than an action target. Concrete answer to “targeted alignment on what?” that What Matters in Learning from Large-Scale Datasets for Robot Manipulation MimicLabs made the empirical case for. Reports better retrieval-relevance vs prior robot-data retrieval methods and higher simulated-robot task success from RoboTok-guided training; specific baselines and numbers not disclosed in the abstract. Open-source (code + HF data/models + project page).
• [2026-09-04] GraspGen-X: Cross-Embodiment 6-DOF Diffusion-based Grasping: GraspGen-X (NVLabs; Han, Chao, Coumans, Eppner, Sundaralingam, Deng, Birchfield, Murali) applies the “hand-designed low-dim shared interface beats learned high-dim features” recipe on the gripper description axis rather than human-to-robot: a 12-dim Swept-Volume vector (two axis-aligned boxes describing space swept by fingers at open + half-open) fed through an MLP, encoding gripper morphology + closing dynamics. Beats learned TSDF (+25% AUC) and learned PointNet++ (+51%) at cross-embodiment grasp generation on 10 held-out real grippers, and gets non-trivial zero-shot performance on 5-finger hands never in the training taxonomy (Surge 0.404, Inspire 0.363). Sibling in position to Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s 3-DoF wrist reduction (low-dim interface on the action side) and C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video‘s object-side contact topology — but the shared interface here is space where contact happens in the gripper base frame rather than object-surface contact patterns. Substrate side: 32 procedurally generated grippers × 8000+ objects × 2 B+ Isaac-simulated grasps beats training on 10 real grippers for every encoder tested, since real grippers don’t cover the target Swept-Volume distribution. Open-source (Apache-2.0 code + NVIDIA Open Model License weights) with a browser-based wizard for onboarding any URDF-defined gripper.
• [2026-09-04] Aria Digital Twin to MCAP — Load ADT Episodes in FiftyOne: Voxel51 (Harpreet Sahota) publishes a ~300-line recipe for converting Meta’s Aria Digital Twin (VRS + 8 ground-truth zips, 12.4 GB per 145 s episode) into a single 2.5 GB MCAP file with 12 synchronized Foxglove-schema channels (RGB, depth, seg, gaze, 3D+2D boxes, skeletons, IMU, 6-DoF pose, semi-dense cloud) that FiftyOne’s multimodal viewer opens natively. Middleware substrate for using ADT — one of the higher-fidelity mocap-grade egocentric HOI corpora used by retargeting recipes on this page — without a per-project loader; documents the tacit sync/axis-convention/static-vs-dynamic-box gotchas that usually eat a week.
• [2026-09-03] Getting a Grip on Robotic Data Collection — Co-Designed Koala Gripper for Force-Mediated Manipulation: RAI Institute’s Koala gripper adds a co-design variant of the make-collection-tool-embodiment-invariant recipe: iterate the human-operated handheld device and the robot-mounted actuated gripper as a single unified platform sharing linkages, finger geometry, kinematics, tracking camera placement, and fingertip force sensors, with a monolithic pivoting double thumb + two underactuated linkage fingers = 3 controllable DoF targeting force-mediated assembly and tool-use tasks. Sibling to mimic (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1)) and XR-1 (Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training) on the hardware-first lineage but distinguished by explicit ergonomic co-design (three prototype sizes, force ratio <3:1 for extended sessions) framed as a data-quality — not just comfort — issue. Announcement post; several hundred demonstrations reported to yield successful policies with no per-task numbers disclosed.
• [2026-09-02] TwinDEX: A Scalable Path from Robot-Free Data Collection to Real-World Dexterous Manipulation: TwinDEX (X Square Robot) pushes the mimic/HandUMI “collection-tool embodiment invariance” recipe up from parallel-gripper and single-hand to a bimanual three-finger nine-DoF dexterous configuration — kinematics, contact surfaces, appearance, and sensor placement mirrored across a wearable and the deployment end-effector. Stakes the position on zero on-robot training or intervention data for a 20+ step chemistry task (tool switches + fine force + bimanual), with 5.3× effective throughput vs on-robot teleop in the multi-task eval. Sharper zero-on-robot-data claim than mimic (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1)) or XR-1 (Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training) which still keep robot teleop in post-training, and a hardware-substrate counterpart to X Square’s own Robots Acquire Manipulation Skills in Seconds from a Single Human Video (HOST) HOST recipe. Announcement surface — no policy architecture, action space, per-task success rates, or per-step ablation disclosed.
• [2026-09-01] What Matters in Learning from Large-Scale Datasets for Robot Manipulation: MimicLabs quantifies the motion-primitive dimension for cross-source robot data: a single-motion co-training set can drop target success by 20% when misaligned, but increasing motion diversity monotonically improves transfer and removes the alignment requirement. Sharpens the “which axes of the retargeted trajectory space must match the deployment target” question for pipelines that mine motions from human video, exoskeletons, or heterogeneous teleop.
• [2026-09-01] The Imitator Game: Benchmarking Robot Imitative Ability Beyond Action Prediction: The Imitator Game (Zhou et al.) operationalizes the interface-location debate this page tracks as a level-graded evaluation axis: L0–L3 progressively widens the demonstration-to-robot-scene gap, and 9 SOTA models are stable L0→L2 then collapse at L3 on functional substitution (same-intent-different-affordance). Human-video-conditioned policies beat caption-conditioned ones under matched protocol — a direct empirical data point in favor of raw sensorimotor stream over language as the human→robot interface. IG-10K (20,000+ paired episodes, 50+ tasks, real+sim) is now the largest environment-aligned paired human-robot substrate on the page and the natural evaluation harness for every “shared interface” position filed here (LUCID’s task-intent, µ₀’s 3D interaction traces, Translation-as-Bridging-Action’s 3-DoF wrist, Hydra-0’s action flow, WAM-TTT’s KV memory, etc.).
• [2026-08-29] Lerrel Pinto — in-context learning for robots is not that hard: Pinto tweet gestures at ICL-as-retargeting-substitute working “out of the box,” adding a claim (no evidence) on the same axis GEN-1.5 and S1 stake out.
• [2026-08-28] OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence: DM0.5 (Dexmal) supplies a “no retargeting layer, put the human demonstration in the context window” datapoint: a start-of-episode human demonstration is observed by the robot within the same ~60 s visual context window that the robot’s own execution then consumes, with in-episode rule-following (e.g. preserving a battery-placement rule the human demonstrated) as the reported outcome. Sibling in position to Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization Zero-WAM’s “in-context prompt as interface” framing but staying inside a single VLA context window rather than a separate world-action model.
• [2026-08-27] Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization: Zero-WAM (Robbyant / Ant Group; Zhou et al.) stakes out an in-context-prompt-as-interface position: rather than retargeting motion, action-representation, task intent, or WFM memory state, keep the raw human demonstration video as an in-context conditioning stream and force the policy to attend to it via an in-context future chunk prediction (IFP) objective. Sibling to GEN-1.5: Embodied Foundation Models are One-Shot Learners GEN-1.5 on the “no explicit retargeting step” axis but distinct in staking the position on a designed training objective + synthetic data pipeline rather than emergence. The load-bearing data-side contribution is HumanGen: 74.2K human-robot ICL pairs across 8.6K tasks generated by rewriting sampled robot trajectories into semantically matched first-person human videos via VLM prompting + Nano Banana / Qwen-Image + two-stage quality gating — a rare instance of synthetic human data (most substrates on this page generate robot data). Reports 46.95% vs 17.45% (LingBot-VA) on 7 held-out RoboTwin 2.0 tasks and real-hardware zero-shot on long-horizon sequential + precision insertion regimes that language alone under-specifies.
• [2026-08-27] Behavior Prompting Policy: Demonstrations as Prompts for Manipulation: BPP (Patel, Pekarek, Castro Hernandez, Song; Stanford REAL Lab; arxiv:2606.30457) adds a mechanistically-specified ICL-substitute-for-retargeting position: keep the raw sensorimotor human demonstration in-context (observation + proprioception + full-rate actions), rely on cross-attention against the current observation to resolve spatial differences via attention-pooled per-timestep chunks. Sibling to GEN-1.5 and S1 on the “no explicit retargeting step” position but distinct in disclaiming the emergence framing — BPP is trained on demonstration-conditioned episodes of the same embodiment at different environment configurations (not human-hand→robot-hand), so the demonstrated retargeting is spatial-configuration adaptation within an embodiment rather than cross-embodiment bridging. Ships iPhUMI, an iPhone 15 Pro + ARKit extension of the UMI gripper with wireless prompt streaming — the same handheld device that collects training data specifies deployment prompts at test time, tightening the collection-tool-embodiment-invariance recipe sibling to Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) and Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training with a same-device dual role.
• [2026-08-27] Project SuperDex — A unified simulation platform for dexterous manipulation: Project SuperDex (Meta) exposes a declarative .superdex_bot library that includes both robot hands and human hands as loadable articulations, with Studio’s bot-mod pipeline for swapping fingertips / hands / arms — the target-embodiment side of the retargeting pipeline, packaged as a first-class authoring workflow rather than a per-project scaffold.
• [2026-08-26] Isaac 0.5: An Open-Weight Embodied Foundation Model: Isaac 0.5 treats three human/handheld video substrates jointly at the pretraining scaling axis — general action-free video (broad visual coverage), egocentric video (375K h, close to hands and contact), and UMI-style handheld-gripper video (375K h, device motion and timing) — held at a fixed 80:30:30 composition and scaled together against teleoperation. Distinct from Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1’s UMI-only 100K-h bet and from EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data EgoScale’s egocentric-only 20,854-h axis — the 210× video-vs-teleop exchange is measured on the combined mixture, not on any single sub-stream. Egocentric supervision is produced via Perceptron’s own Mk1 VLM (Introducing Perceptron Mk1) with a three-level structured hierarchy (recording / activity segment / per-hand instance).
• [2026-08-25] Introducing Index: Building The World's Largest and Most Diverse Physical Dataset: Figure Index blog details the collection side of the human-to-robot pipeline — 30 min/s video ingest, embedding-based dedup, and hierarchical captioning — but does not disclose the retargeting step from unstructured phone video into Helix’s action space.
• [2026-08-25] Figure launches Index — crowdsourced app-based real-world video dataset for humanoid robots: Figure launches Index — an app-based paid-crowdsourced pipeline for collecting real-world human video (unstructured phone-recorded footage across 108 countries); retargeting details undisclosed but the coarse collection format will force the interface-location question at the largest funded humanoid company.
• [2026-08-24] Compute-Optimal Scaling Laws for Human Motion Generation: Cartwheel releases a ~12,000-hour (5.56M-clip) motion corpus assembled by running a 3D-reconstruction + physical-QA pipeline (“Comic”) over a ~423k-hour raw video catalog with ~35× quality-gate reduction, then reports Chinchilla-style compute-optimal scaling laws on it — a substrate ~4× the next-largest motion corpus, targeted directly at the “cheap human motion → target embodiment” retargeting problem this page tracks. Complements the pipeline side of WARP: Whole-body Retargeting from Offline Human Motion via Closed-Form c-SEW (WARP) and ReForce — force-aware retargeting via closed-loop residual force tracker (ReForce) with the data-side scaling result: at motion-corpus scale, both AR and flow-matching motion generators still improve predictably along Chinchilla exponents, so retargeting-dataset-size choices now have a scaling-law calculus attached rather than being a “collect more and hope” line item.
• [2026-08-21] T-Rex: Tactile-Reactive Dexterous Manipulation: T-Rex reuses the EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data retargeting recipe wholesale — 22,889 h of egocentric human video with retargeted 22-DoF Sharpa-hand joint actions in a unified action space — as the pretraining leg of a three-stage recipe whose second stage now adds tactile mid-training on top of the visuomotor priors. Ablations quantify the retargeting stage independently: from-scratch 18% → +human-pretraining 34% → +tactile-mid-training 45% → full 65% across six benchmark tasks. Concrete evidence that the joint-space 22-DoF retargeting channel EgoScale established is composable with downstream modality additions — the pretrained visuomotor priors do not need to be re-derived when a new sensing modality is grafted on.
• [2026-08-21] Hydra-0: Action Flow for Generalist World Modeling and Control: Hydra-0 (NVIDIA + Brown/Columbia) stakes out a new position on the interface-location axis: the shared human–robot interface is 2D pixel motion (action flow), one abstraction level below μ₀: A Scalable 3D Interaction-Trace World Model‘s 3D interaction traces and UMA: Unified Motion-Action Modeling for Heterogeneous Robot Learning‘s 3D object motion. The inverse mode — object flow transferred from a human demonstration drives the world action model to predict compatible robot motion, with a trained action head decoding to executable actions — is the closest filed sibling to LUCID’s object-flow-as-intent recipe (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition) but keeps the intent-to-action mapping inside a single video-generation backbone rather than delegating to a separate sim-trained sensorimotor policy. Reports r = 0.96 replay-vs-reference success on RoboLab, plus 90.4% / 60.2% lower robot- / object-motion error vs an action-conditioned baseline.
• [2026-08-20] OakInk: A Large-scale Knowledge Repository for Understanding Hand-Object Interaction: OakInk (Yang et al., SJTU MVIG, CVPR 2022) is the contact-first-transfer ancestor of the recipes on this page — 1,800 scanned objects + 50K hand-object interactions built by recording 1K real demos and using Tink (DeepSDF interpolation + explicit contact mapping + iterative pose refinement with SDF interpenetration cost) to transfer them onto 1,700 virtual counterparts of similar affordance. First filed HOI dataset to bundle intent labels, dynamic interaction, human-to-human handover, hand-object contact, and 6-DoF object pose in one corpus (Table 1). Methodological predecessor of C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video‘s object-side contact-topology retargeting interface and one of the two “grasp/object-aware pose” substrate slots named by Yu Xiang — 12 hand-object interaction datasets worth knowing for robot learning.
• [2026-08-20] GEN-0: Embodied Foundation Models That Scale with Physical Interaction: GEN-0 (Generalist AI, Nov 2025) covers 6-DoF, 7-DoF, and 16+-DoF semi-humanoid embodiments from a single base model — cross-embodiment as an architectural property rather than a retargeting adapter, sibling in position to Towards Machines with a Thousand Hands‘s ~9,000-end-effector coverage on the same underlying engine.
• [2026-08-19] GEN-1.5: Embodied Foundation Models are One-Shot Learners: GEN-1.5 adds an “ICL substitute for retargeting” datapoint distinct from every other filing on this page: a human demonstrating a task with their own hands in view of the robot’s cameras is treated as a physical prompt in the 30 s context window, and the robot reproduces the task on its own hands with no gradient updates and no explicit retargeting step. Directly counter to the ReActor / c-SEW / Translation-as-Bridging-Action recipe family that collapses human embodiment onto a canonical robot representation — GEN-1.5 keeps the raw human-hand video and asks the model to bridge the embodiment gap in a single forward pass. Emergent from pretraining scale rather than from a designed cross-embodiment representation; no numbers reported for the human-hand-prompt case specifically (aggregate is 59% avg one-shot across 10 tasks).
• [2026-08-19] WARP: Whole-body Retargeting from Offline Human Motion via Closed-Form c-SEW: WARP (Chen, Kong, Zhang, Yang, Zhu, Kousik, Xu — Georgia Tech / Danfei Xu’s group) reframes the concept’s interface-location axis by arguing the real lever is solver class, not interface location: a closed-form Shoulder–Elbow–Wrist analytical IK (c-SEW) with palm pose as a hard constraint gives exactly-one-robot-config-per-human-pose, cutting palm error >150× vs MINK-EF, hitting machine-precision orientation error, and producing the most self-collision- and joint-limit-feasible trajectories on 514 BONES-SEED clips. Directly targets the offline-replay regime where the retargeted trajectory is the supervision (no operator to absorb error). Cross-embodiment demos on RB-Y1, dual Kinova3, OpenArm; reported as enabling the whole-body-mobile-manipulation policy trained from offline human demos alone. Complements ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting (physics-in-loop bilevel) by fixing hardware-executability upstream at the IK solver rather than downstream in an outer RL loop.
• [2026-08-17] ReForce — force-aware retargeting via closed-loop residual force tracker: ReForce (@wu876768 tweet) stakes out a new position on the interface-location axis: the shared human–robot interface is a force-tracking residual learned in sim, added on top of an existing motion+force reference (teleop or a learned ACT policy), rather than a reduced action space (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), a physics-in-loop bilevel retargeter (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting), or a kinematic-reconstruction pipeline (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos). Reports lowest force-tracking error and 30%→70% force-safe success with 7/10→0/10 missing-contact failures on paper-cup grasping under a learned-ACT reference; qualitative demos of spring pinching, board wiping, tongs squeezing, cup pouring. No linked paper.
• [2026-08-17] Yu Xiang — 12 hand-object interaction datasets worth knowing for robot learning: Yu Xiang (IRVL) posts a curated 12-item directory of hand-object interaction datasets (FreiHAND, RHD, HO-Cap, HO3D/H2O-3D, ObMan, DexYCB, HOI4D, OakInk v1/v2, ARCTIC, PALM, TACO) with the framing claim that hand pose is the load-bearing bridge for learning from human interaction and that the hard part is cross-dataset comparability of pose formats, camera models, coordinate frames, and object annotations — names the substrate that filed retargeting recipes (C2Dex, DO AS I DO, ProxyPose) implicitly compete over.
• [2026-08-15] Simulation Pre-training for Dexterity (SPD): SPD (Kamat, Rashid et al.; Stanford + MIT + Scale AI — Finn / Isola / Karen Liu) stakes out a new position on the interface-location axis: skip the human-to-robot retargeting problem entirely by putting the human demonstrator into a MuJoCo VR headset controlling the exact target embodiment (22-DoF Sharpa hand + YAM Pro arms). Five operators × 75 h × six scenes in one week; no post-hoc retargeting, no kinematic gap, and reset costs go to zero. Complements Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training XR-1’s UMI-scale bet and Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic’s hardware-locked wearable by moving the collection environment into simulation rather than making the collection tool embodiment-invariant. 1–2 h real teleop fine-tune per task beats from-scratch BC by +18 pp average task progress across five real dexterous tasks. Dataset + VR software (spd-75h, spd-vr) + six curated MuJoCo scenes released.
• [2026-08-14] Ego-OSCAR: Egocentric Open source Stereo CAptuRe System: Ego-OSCAR (FPV Labs) sits at the substrate layer of the retargeting recipe debate rather than the interface layer: a fully open-source head-mounted stereo-inertial capture device (<USD 200 BoM, 3D-printed enclosure, hardware-synced global-shutter stereo + 6-axis IMU) plus a 550 h / 100+ environment egocentric dataset with 209,315 free-form action segments and per-frame 3D hand reconstructions. Explicitly positioned against closed platforms like Project Aria — the argument is that the data bottleneck for VLA/WAM pipelines is now the capture device, not the model, and closed rigs cannot be reproduced across distributed contributor networks. Closest sibling to HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone HiFi-UMI but targets pure egocentric video+IMU rather than bimanual UMI action fidelity, and unlike HOT3D (HOT3D — Egocentric Multi-View Dataset for 3D Hand-Object Interaction Understanding) uses consumer-grade parts + offline pose (VINS-Fusion converges on 12/20 held-out sequences vs 15/20 for a RealSense) to enable crowdsourcing.
• [2026-08-14] H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models: H2R-Bench (Rong, Shi, Ma, Cao et al. — SJTU + collaborators) gives this cluster its first dedicated video-generative retargeting benchmark: 11 SOTA video generation models scored on transforming egocentric human demos into robot-centric videos across 6 manipulation families × 2 embodiments, with five source-grounded axes (goal-state completion, action-event completion, functional contact transfer, embodiment correctness, general video quality). Complements the training-recipe entries on the page (kinematic retargeting, reduced-DoF bridging actions, physics-in-loop optimization, WFM-memory approaches) with a diagnostic framework for the video-model substrate itself, and reports that leading video WMs currently fail on embodiment consistency, functional interaction, and task execution.
• [2026-08-13] HOT3D — Egocentric Multi-View Dataset for 3D Hand-Object Interaction Understanding: HOT3D (Meta Reality Labs / Project Aria) provides the accurate mocap-grade 3D hand + object ground truth on real egocentric multi-view streams (Aria + Quest 3, 33 PBR-scanned objects, gaze) that dexterous kinematic-retargeting recipes (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos, HUG: Human Universal Grasping) currently source from smaller or noisier corpora — a supervision-quality-focused substrate to sit alongside the scale-focused ones (EgoScale, UniHand, HumanNet) already on this page.
• [2026-08-12] C2Dex: Contact-Consistent Reconstruction and Retargeting for Dexterous Manipulation from Monocular Video: C2Dex (Nanjing / TUM) stakes out a ninth position on the interface-location axis: the shared embodiment-invariant interface is object-side contact topology in canonical object space, aggregated from noisy per-frame HOI reconstructions to serve as both a trajectory-level constraint on human HOI reconstruction and a transfer target for Laplacian interaction optimization onto the dexterous robot hand, with residual RL closing the physics gap. Sibling to CHORD (CHORD: Object-Centric Contact Wrench Guidance for Dexterous Manipulation) at the geometric rather than force/wrench level. Reports 57.78% vs 17.78% end-to-end trajectory success on DexYCB and 26.67% vs 10.00% on TACO over strongest baselines under matched evaluation — the largest reported margin over kinematic-only retargeting recipes (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos) filed on this page.
• [2026-08-10] Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models: Dyna-2 stakes out the minimal-interface + world-modeling + scale endpoint of the interface-location axis: only wrist EE trajectory + continuous grasp signal (thumb–index aperture) from a 3D hand-pose pipeline, no visual/kinematic bridging, paid for with 1M+ hours of egocentric human video and joint video+action denoising. Directly contradicts the EgoScale finding that 22-DoF joint-space retargeting is the winning channel — Dyna-2 argues action-only supervision loses at every scale on 39/39 held-out YAM tasks, and the transfer comes from world modeling, not action-representation richness. Cross-embodiment offline scaling shows an emergence-like inflection between 10k and 100k h.
• [2026-08-10] VLAff: Vision-Language-Affordance Model for Unified Actionable Affordances: VLAff (JSK Lab, U-Tokyo) stakes out a new interface position — retarget an affordance triple (visual heatmap + 6-DoF grasp pose + trajectory waypoints) from egocentric human video via SfM + hand-mesh reconstruction, rather than motion (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), reduced action space (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), task intent (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition), or task progress (Robots Acquire Manipulation Skills in Seconds from a Single Human Video (HOST) HOST). Auto-extracted EgoAffordance corpus of 204K episodes / 5.6M visual + 11.6M grasp+trajectory affordances drives a unified VLM foundation model reporting SOTA visual-affordance prediction and zero-shot real-robot manipulation via 3D-scene-grounded execution.
• [2026-08-10] SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation: SimToolReal is the object-frame-goal-pose variant of functional retargeting: instead of transferring human hand motion (kinematic) or task outcomes with per-task RL, it extracts a single 6-D object-pose trajectory from an RGB-D human video and lets a general goal-conditioned RL policy discover the underlying dexterous skills (in-hand rotation, stable grasp) zero-shot. Beats kinematic-retargeting baselines by +37% Task Progress on DexToolBench.
• [2026-08-08] MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives: MotionBricks (NVIDIA, TOG 2026) inverts the concept’s usual substrate direction: instead of retargeting human video onto a robot, it deploys a character-animation model (trained on mocap + animation data at 350k clips) zero-shot as a Unitree G1 humanoid controller — no paired human-robot training stage, no per-embodiment fine-tuning, no task-specific tagging. Adds a new position to the interface-location axis: the shared backbone lives at the motion-token layer with a keyframe-command interface, and the same frozen model serves both UE5 character animation and real-time humanoid control. Distinct from ScaleBFM (Scaling Behavior Foundation Model for Humanoid Robots) which uses PPO on retargeted mocap and from ReActor (ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting) which puts physics inside the outer retargeting loop.
• [2026-08-08] AI Sapiens Open Source — ROBOTIS Docs Directory: ROBOTIS’s AI Sapiens docs page documents soma-retargeter (forked from NVIDIA/soma-retargeter) and kimodo (forked from nv-tlabs/kimodo) as the K1 humanoid’s retargeting pipeline components, alongside GEM-X for video-to-motion generation.
• [2026-08-06] RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data: RekaDaily-10k (Reka, Apache 2.0) adds a 10,312 h unscripted-household open pretraining substrate for the human-video-to-robot retargeting recipes on this page — roughly half the size of EgoScale’s 20,854 h and complementary in scope (unscripted, multi-region, household, per-clip captions).
• [2026-08-04] Assemble Bench and CG-DAgger — NIST small-parts assembly benchmark on Isaac Lab Arena + code-guided DAgger with zero teleop: HUD explicitly rejects teleoperation-derived demos for its 14 contact-rich DROID tasks, arguing the alignment tolerance is too fine for human-piloted trajectories to be clean supervision — a counterpoint to the growing “scale human teleop” thesis, favoring scripted-expert data for hard contact tasks.
• [2026-08-03] Robots Acquire Manipulation Skills in Seconds from a Single Human Video (HOST): HOST (X Square Robot + Tsinghua + BIT) stakes out a distinct interface position: retarget a shared task-progress manifold coordinate rather than motion, action representation, task intent, or WFM memory state — then predict the robot’s own future observations from that coordinate and derive actions. Frozen parameters at acquisition time (no updates), ~29 s per new skill from a single human video, 62% average success on 8-task baseline (+45 pp over strongest no-update method) across 50 novel tasks — and exceeds a 50-demo/task fine-tuned baseline at 50× fewer demos and 507× faster per skill. Sibling to WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time on the frozen-parameter one-shot axis but with the human-robot correspondence in a progress manifold rather than KV memory, and no paired human-robot meta-training stage required.
• [2026-08-03] Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand: Aero Hand Open offers a $314 open-source anthropomorphic 16-joint hand — a low-cost target embodiment that lowers the hardware barrier for human-video-to-robot retargeting work outside well-funded labs.
• [2026-07-31] PhiZero: A World Model Built Around Physical Language: PhiZero demonstrates zero-shot Human → G1 humanoid and Human hand → Sharpa dexterous hand retargeting by reusing a source clip’s predicted physical-language sequence with a new first frame — no paired human-robot data or embodiment-specific retraining of the reasoner.
• [2026-07-29] HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone: HiFi-UMI (Simple AI / Simple World Lab) sharpens the “make-the-collection-tool-embodiment-invariant” position by trading corpus scale for rig fidelity: 3 mm workspace-local end-effector accuracy via head-mounted offline stereo-inertial SLAM, <40 μs cross-sensor GPIO sync, native (not reconstructed) inter-gripper relative pose, and six ultra-wide cameras per bimanual demo. The load-bearing empirical claim is stronger than Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training‘s scale-first UMI recipe (which still keeps 7,200 h teleop in post-training): a policy post-trained solely on HiFi-UMI data — no real-robot anchor — matches an in-domain teleop baseline across three separate backbones (StarVLA-QwenPI −2.5, OpenPI-π0.5 +3.1, LingBot-VA −0.6 pp) and reaches 85% on a precision insertion task with zero in-scene HiFi-UMI trajectories. 2,000-hour HiFi-UMI-2K corpus released under CC-BY-4.0.
• [2026-07-28] Data Pyramid for Embodied Manipulation: Data Pyramid names cross-embodiment action alignment as one of six open challenges for embodied-manipulation data recipes, treating it as a data-side lever rather than an interface-design problem — sits alongside the interface-location debate this concept page tracks (physics-in-loop retargeting / 4D reconstruction / constraint-RL / reduced-DoF / hardware-locked kinematic equivalence / shared latent) with the meta-position that the field lacks principled recipes for mixing egocentric, UMI, and teleop sources across embodiments.
• [2026-07-28] ModPack: An Extensible Teleoperation Interface for Bimanual Mobile Manipulation: ModPack (Citron, Zbizika, Liu, Song / Stanford REALab) stakes out a new position on the interface-location axis: fix the embodiment gap at the operator rig rather than at the collection-tool geometry (Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1), HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet) HandUMI) or the action-representation (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots) — a shared 3D-printed backpack (compute + power) plus GELLO-style leader arms kinematically equivalent to each follower (ARX5-based bimanual, RB-Y1m) plus plug-and-play modules (haptic feedback, active perception, mobile manipulation). Fully open-sourced hardware + software (Apache-style, with BOM). Sibling to RAIDEN: A Toolkit for Policy Learning with YAM Bimanual Robot Arms RAIDEN’s software-toolkit take on the same open-teleop-substrate need.
• [2026-07-27] Is Diversity All You Need for Scalable Robotic Manipulation?: Shi et al.’s embodiment-diversity result adds a datapoint to the “which embodiments should appear in the pretraining substrate” question: on ManiSkill (Franka), RoboTwin (Arx), and real Agilex (Piper), RDT pre-trained solely on AgiBot G1 (RDT-AWB) matches or beats RDT pre-trained on OXE (which includes Franka), with a favorable power-law slope in fine-tuning data. Sharpens the case that high-quality single-embodiment pre-training is a viable substitute for multi-embodiment aggregation and reframes the retargeting problem as one solved by fine-tuning rather than by pretraining data composition. Also introduces a temporal retargeting axis — the Velocity Model debiases demonstrator-speed variation independently of demonstrator spatial habits — complementing the kinematic-only retargeting recipes already on the page.
• [2026-07-27] RW-RL Dataset: Real-World Reinforcement Learning for Robots: RW-RL Dataset captures human involvement not as source-domain data to retarget onto a robot but as in-loop intervention on the robot itself — human teleoperation for the imitation prior, human takeover for recovery from off-distribution states with intervention start/end frames + structured intervention reasons labeled per episode, and autonomous rollouts (successful + failed) as the third channel. Complements the human-video-to-robot retargeting lineage with an on-robot answer for the same underlying failure mode (policy drifts off the demonstration manifold), targeting 1000+ hours across 4 robot series.
• [2026-07-25] Light Origins tweet 2080861398712946779 (content not retrievable at filing time): Light Origins (Singapore/Shenzhen/Beijing, founded by ChatGPT co-inventor Roger Jiang) frames the whole company thesis as “compress human behavior into Universal Action Representations” transferable across morphologies — another vote for a single shared behavior manifold rather than per-embodiment interfaces.
• [2026-07-23] EgoRecovery: Acquiring Failure Recovery Ability Through Human Recovery Demonstration: EgoRecovery (Ge et al., Fudan / Yu-Gang Jiang) stakes out a distinct interface position: retarget a corrective intent (timing + magnitude of correction) rather than motion, action-representation, or generic task intent — and gate it on only when a learned recovery-state predictor fires. This narrow scoping is what makes the small-robot-data recipe plausible: the shared human–robot latent only has to represent when-and-how-strongly to correct, not the full motion manifold. Complements LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition (task-level scene flow) by narrowing the intent abstraction to recovery states, and gives a data-side counterpart to HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining with a claimed >10× per-hour egocentric-vs-teleop recovery-data rate.
• [2026-07-23] Towards Machines with a Thousand Hands: GEN-1’s “Thousand Hands” position is the sharpest counter-recipe on this page’s interface-location axis: don’t collapse the demonstrator or collection tool onto a canonical robot embodiment at all — instead scale the robot side to ~9,000 end-effector variations across >500K hours and let the model treat each hand as a different “language” for physical interaction. Directly contrasts Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training‘s embodiment-free-UMI recipe (which fixes one collection embodiment across scenarios) and Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1)‘s hardware-locked kinematic equivalence (which fixes one hand across collector and robot) by moving the fix onto the model’s multi-embodiment capacity rather than onto the data pipeline. Task-vector analysis of fine-tuning updates gives the first filed subsystem-localized measurement of how novel each embodiment is to a pretrained policy, sharpening the retargeting-quality vs retargeting-diversity open question with a concrete measurement primitive.
• [2026-07-20] Danfei Xu — two root-level paradigm shifts in robot learning: Sim2Real for locomotion, Behavior Cloning for manipulation: Danfei Xu’s “BC-for-manipulation as the paradigm shift” framing is the implicit motivation for the retargeting subfield — extracting demonstrations from human video is a direct response to BC’s data-hunger.
• [2026-07-20] Scaling Behavior Foundation Model for Humanoid Robots: ScaleBFM (Zeng et al.) treats retargeted human motion as the pretraining substrate for a humanoid BFM rather than as a per-clip control target: 102M frames aggregated across multiple open-source motion datasets and retargeted onto the target humanoid, with PPO-based motion tracking as the proxy task. The load-bearing finding for this concept: reference-motion scaling splits into two regimes — homogeneous (more in-domain data, similar behavior coverage) yields only marginal gains, while heterogeneous (diverse sources expanding behavior coverage) yields substantial OOD gains. Contrasts with ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting‘s physics-in-loop retargeting-quality lever by showing source-diversity is a distinct lever that matters even under a fixed kinematic retargeter; sharpens the “retargeting quality vs retargeting diversity” open question on the page with a quantitative answer favoring diversity for OOD generalization.
• [2026-07-18] HandUMI — Robot-free bimanual data collection retargetable to any parallel-gripper arm (Brik H. Meza launch tweet): HandUMI (Brik H. Meza et al.) packages the make-collection-tool-embodiment-invariant position as an open-source, Apache-2.0 hardware+software stack: robot-free wearable capture that replays onto four distinct bimanual parallel-gripper arms (Agilex PiPER, OpenArm, TRLC-DK1, I2RT YAM) by swapping the gripper module, with LeRobot v3-compatible output, built-in calibration/QA, and both sim replay and real teleoperation supported. Complements Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training‘s at-scale UMI-based pretraining evidence by making the same collection substrate broadly available to small teams and across multiple embodiments rather than a single vendor rig.
• [2026-07-18] Towards Predictive, Aligned, and Scalable Robot Learning (Lumo-2): Lumo-2 (Astribot) headlines “Human-to-Robot learning” as one of three flagship capabilities alongside predictive world dynamics and long-horizon manipulation; the launch showcase demonstrates 20+ household tasks on a real humanoid with cross-embodiment learning from human video contributing to the training substrate.
• [2026-07-17] WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time: WAM-TTT (Feng, Lyu, Wang et al.) proposes a distinct position on the interface-location axis: don’t retarget human motion to robot embodiment at all, and don’t retarget task intent either — instead absorb the raw human video into a frozen WAM’s key–value memory via self-supervised video prediction, and rely on a paired-human-robot meta-training stage with a KV reconstruction objective to align memory content with robot-behavior-relevant features. The human-robot correspondence lives inside the frozen model’s internal state rather than in an explicit action-space reduction (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), 4D kinematic reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), or intent abstraction (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition). At deployment only unlabeled human videos are needed; no robot demos, no annotations, no fine-tuning. Reports consistent wins over in-context human-video conditioning baselines across manipulation and generalization settings.
• [2026-07-17] Mixture of Frames Policy: Multi-Frame Action Denoising for Bimanual Mobile Manipulation: Mixture of Frames Policy (Wang, Song, Bohg et al.) reframes the “which frame to represent motion in” problem — a recurring headache for cross-embodiment retargeting — as an MoE routing decision inside a diffusion policy: denoise the bimanual action chunk in base / left-hand / right-hand / rel-traj frames in parallel and learn per-subtask expert weights. Measures a 15.0-point gap between best- and worst-per-task frame choice and shows learned per-subtask routing beats even the oracle-per-task frame choice (66.8 vs 63.8), suggesting the frame problem is at least partly architectural rather than purely a data/kinematics engineering concern.
• [2026-07-16] Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1): mimic robotics stakes out a hardware-first position on the interface-location axis: fix the embodiment-invariant interface at the hand-kinematics level by shipping a wearable exoskeleton (U1) whose rigid linkage mechanically constrains the human hand into the motion space of the robot hand (M1), eliminating retargeting at collection time. This is the same shortcut as Xiaomi-Robotics-1 (XR-1) — Scaling VLA Foundation Models with 100K Hours of Embodiment-Free UMI Pre-training‘s UMI-based collection but extended one embodiment upward from two-finger grippers to a fully dexterous 15-DoF anthropomorphic hand; complements the algorithmic positions (physics-in-loop ReActor, reduced-DoF Translation-as-Bridging-Action, task-intent LUCID) by moving the fix into the collection hardware itself. Tactile + joint-encoder + wrist-camera sensing is mirrored 1:1 across M1 and U1 so demonstrator observations match robot observations exactly.
• [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) proposes an eighth position on the interface-location axis: make the collection tool itself embodiment-invariant — hand-held UMI grippers are the retargeting target at collection time, so the kinematic + contact-mechanics gap between demonstrator and robot is closed at data-gathering, not by post-hoc reconstruction (Do as I Do: Dexterous Manipulation Data from Everyday Human Videos), reduction (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots), or intent abstraction (LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition). The recipe’s headline evidence is that this shortcut scales cleanly: 100,000 hours of UMI pre-training across 1,700+ scenarios yields smooth pre-train → post-train real-robot success curves with no saturation, and 75% vs π0.5’s 40% at <10 h/task of downstream robot data. Sits alongside EgoScale’s 20,854 h egocentric-video result as one of the two filed data points supporting scale-as-the-lever on non-teleop substrates.
• [2026-07-13] HUG: Human Universal Grasping: HUG (NYU / Pinto group) stakes out a seventh position on the interface-design axis: keep the full 51-DoF MANO articulation, normalize hand shape to a canonical β (so θ denotes the same grasp across collectors), and pay for retargeting with a high-quality capture stream — 1M-HUGs, 1M egocentric Aria-Gen-2 RGB-D frames from 6,707 in-the-wild human grasps across 41 buildings + camera-pose back-propagation for view multiplication. Reaches 73% tabletop and 62% in-the-wild real-world grasp success on 30 unseen HUG-Bench objects, beating sim-trained Dex1B (44%) and parallel-jaw CAP (33%), with zero-shot deployment across Ability, WUJI, and Aria/ZED cameras. Sits between Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s reduce-to-3-DoF and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data‘s scale-with-22-DoF-joints — keeps full articulation but with 750× less pretraining data than EgoScale.
• [2026-07-11] Cross-Embodiment Robot Manipulation via a Unified Hand Action Space: UHAS (UT Dallas / Casas, Xiang) proposes a sixth position on the interface-location axis: a geometric fingertip-surface interface — a canonical sphere with lateral rotation and radial displacement parameters — shared across Allegro / LEAP / Shadow / MANO Human hands, with a Cascade IK solver mapping the shared sphere back to per-embodiment joint targets. A single RL policy trained in the sphere-deformation space controls all four hands on in-hand cube reorientation and transfers zero-shot to unseen embodiments; real-world deployment on LEAP and Allegro from the same policy. Contrasts most directly with Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots (drop rotation) and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data (keep 22-DoF joints + scale) by reparameterizing rotation into a geometry that is embodiment-invariant by construction.
• [2026-07-11] LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition: LUCID (CMU; Gupta, Shi, Yuan) proposes a sixth lever: don’t retarget human motion at all — instead retarget task intent (short-horizon 3D object flow + palm-pose reference) and delegate morphology-specific execution to a sim-trained sensorimotor policy. The same intent model drives a LEAP dexterous hand and a parallel-jaw gripper at matching 19/30 success on push-T + cable routing supervised by 1 hr of smartphone video, and closed-loop intent re-querying beats a Veo-3.1 open-loop planner 73% vs 28% on stirring/wiping/binning. Structurally distinct from the motion-retargeting recipes on this page: the interface between human data and robot control is a scene-level prediction (what should happen next) rather than a joint-level or end-effector-level trajectory, which is why one intent module trivially spans two very different embodiments.
• [2026-07-10] LingBot-VA 2.0: Native Video-Action Pretraining for Generalizable Robot Control: Human demonstration videos are used as in-context visual task prompts (§2.3.4) to the causal video-action policy, and human-robot co-training brings egocentric human videos into the shared world model (§2.3.5). Complements the reduction/filtering positions (HumanScale filtering; Translation-as-a-Bridging-Action retargeting) with a shared-latent position: the human video is consumed at the same tokenizer layer as robot video, and the latent-action tokenizer’s bottleneck yields action-relevant supervision from human video without pseudo-action labels.
• [2026-07-09] From Foundation to Application: Improving VLA Models in Practice (LingBot-VLA 2.0): LingBot-VLA 2.0 takes the keep-both-channels-at-scale position: 10,000 hours of egocentric human video are mixed directly into the same pretraining stage as 50,000 hours of robot trajectories spanning 20 configurations, rather than as a separate mid-training phase (ACE-Ego-0), a filtered auxiliary channel (HumanScale), or a reduced 3-DoF bridging interface (Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots). Provides a corpus-scale reference point ~50% the size of EgoScale’s egocentric-only 20,854 hours but combined with 2.4× more robot data.
• [2026-07-02] HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation: HABIT (KAIST; Song et al.) sits adjacent to but distinct from this concept: it does not retarget human motion to a robot embodiment (the robot is teleoperated), but it does retarget the coordination interface — every trigger cue in the recorded data must exist in the camera stream, and Supervisor tasks force the robot to infer intent from human gesture rather than shortcutting to workflow knowledge. Complements the motion-retargeting recipes on this page by showing that on the interaction-cue channel, the analogous “drop the unobservable signals” principle likewise produces policies that transfer.
• [2026-06-30] TrueLabel — Physical AI Dataset Directory: Robotics, Humanoid & Egocentric Data: TrueLabel Physical AI Dataset Directory provides a side-by-side license-and-scale view across the egocentric corpus (Ego4D, EPIC-KITCHENS, HOI4D, DexYCB, xperience-10m, 10Kh-RealOmin-OpenData, UniHand_Preview) and the teleop corpus (AgiBot World, DROID, LeRobot releases) that this concept’s filed papers debate as pretraining substrates.
• [2026-06-30] EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data: EgoScale (NVIDIA GEAR) is a sixth lever on the action-channel-design axis: keep the full 22-DoF retargeted Sharpa-hand joint actions plus relative wrist motion, and pay for the noise with raw pretraining scale (20,854 h of egocentric video). Directly contradicts the Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots drop-the-noisy-channels recipe by ablating wrist-only and fingertip-based representations and finding the 22-DoF joint channel wins across all 5 contact-sensitive tasks. Cross-embodiment transfer to a Unitree G1 7-DoF tri-finger hand (+30%+ absolute) is the first filed evidence that a high-DoF retargeting target functions as a reusable motor prior across substantially different hand kinematics.
• [2026-06-30] datasets.bot — Curated catalog of robotics and embodied-AI training datasets: datasets.bot — third-party browseable catalog of robotics/embodied-AI training datasets (LeRobot, AgiBot World, HIW-500, GR00T X-Embodiment Sim, FTP-1, ABC-130k, DreamDojo, …) with per-dataset modality tags (rgb/depth/ee_pose/proprioception/tactile/force_torque/ego), license, and popularity proxies; useful as a scoping aid for “what human/robot data exists under what license” before picking a retargeting substrate.
• [2026-06-30] ACE-Ego-0: Unifying Egocentric Human and Robotic Data for VLA Pretraining: ACE-Ego-0 absorbs the human-robot embodiment gap inside the training objective rather than at the representation or physics level: keeps the full action space, unifies streams via camera-space actions + morphology conditioning + time-aligned chunking, and pays for the resulting pseudo-action label noise with a reliability-aware loss and human auxiliary loss — directly contrasting Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots‘s drop-the-noisy-channels recipe on the same problem.
• [2026-06-30] Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots: Fourth distinct lever for the cluster — action-representation design rather than physics/perception/RL-constraint. Proposes a 3-DoF bridging action (relative wrist translation in head-camera frame, 30-step chunk) shared by humans and robots, plus interleaved-action-token + attention-masking training that absorbs three data sources with different action coverage. Outperforms 6-DoF baseline from scratch on the ByteMini bi-manual suite (44.58 vs 34.67% task progress, 22.50 vs 12.50% success rate) and shows that adding large-scale human-action pretraining strongly outperforms pick-and-place-only training.
• [2026-06-19] ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control: ConTrack (UCSD; Xiaolong Wang group, ECCV 2026) — dual-variable RL update reallocates authority between object-tracking constraint and hand-style fidelity online, eliminating per-clip reward tuning. Closed-loop-control half of the same dexterous-hand pipeline that Do as I Do attacks at the kinematic-reconstruction level.
• [2026-06-18] Do as I Do: Dexterous Manipulation Data from Everyday Human Videos: DO AS I DO (Malik group; Berkeley + Meta) — kinematic-only pipeline: 4D hand-object reconstruction from monocular RGB + retargeting to multi-fingered robot hand, end-to-end. Targets in-the-wild data the lab-captured corpora can’t reach; positions itself as Malik group’s positive answer to its own sensorimotor counter-position.
• [2026-05-24] ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting: ReActor (Disney Research Switzerland) — bilevel optimization with physics in the outer retargeting loop and an RL tracking policy in the inner loop. Targets full-body humanoids + a quadruped (Disney animatronics), validated on hardware. The framework derives an approximate gradient on the outer loss so the joint optimization doesn’t require unrolling through the RL inner loop.

Open questions

• At what level should the embodiment-invariant interface live — joint pose with adaptive correspondences (ReActor), 4D hand-object reconstruction with kinematic retargeting (Do as I Do), object trajectory as a constraint (ConTrack), or a deliberately reduced action space that drops the noisy channels (Translation-as-Bridging-Action)?
• Do the physics-in-loop and action-representation-reduction recipes compose — i.e. does dropping wrist rotation on the human side help even when a downstream physics-simulator-grounded policy is solving the dynamics? No filed paper combines them.
• Does the full-body humanoid retargeting problem (ReActor) want a different interface than the dexterous-hand retargeting problem (ConTrack / Do as I Do), or is there a unified parameterization?
• How well does the bridging-action / interleaved-token recipe generalize to full bi-manual dexterous tasks (not just parallel-gripper bi-manual ByteMini tasks where rotation can be cleanly discarded)?
• ReActor and Do as I Do both bypass the teleop vs human-video substrate debate by treating human motion as a data source and retargeting as the recipe — does this make HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining‘s matched-volume head-to-head measurement moot, or does retargeting quality become the new lever?

Papers

2026-06

• Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots — Translation as a Bridging Action: Transferring Manipulation Skills from Humans to Robots (published 2026-06-25)
• Do as I Do: Dexterous Manipulation Data from Everyday Human Videos — Do as I Do: Dexterous Manipulation Data from Everyday Human Videos (published 2026-06-17)
• ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control — ConTrack: Constrained Hand Motion Tracking with Adaptive Trade-off Control (published 2026-06-02)

2026-05

• ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting — ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting (published 2026-05-07)