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

• [2026-08-29] 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)