Pith. sign in

REVIEW 3 major objections 6 minor 18 references

RoboReact: Agentic Skill Distillation from Generated Egocentric Videos for Generalizable Whole-Body Manipulation

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RoboReact claims that a single egocentric RGB-D image plus a language instruction can yield an executable whole-body humanoid skill, with an 81.3% mean terminal success rate across four real-world bimanual tasks.

desk verdict A genuinely new pipeline for distilling whole-body humanoid skills from generated video, with real-robot evidence that is plausible but statistically thin and one unablated human-hint confound. read the letter →

arxiv 2608.03387 v2 pith:CCFLPONX submitted 2026-08-04 cs.RO

Shuliang He , Shuai Wang , Bo Yue , Junchi Teng , Changyu Wang , Guiliang Liu This is my paper · ORCID
classification cs.RO
keywords humanoidmanipulationskilldistillationvideogenerationvision-languagemodelagenticrefinementwhole-bodycontrolegocentricRGB-Dobject-centricre-grounding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that humanoid robots can acquire dexterous whole-body manipulation skills without teleoperation, human demonstration data, or task-specific reward engineering. The input is one egocentric RGB-D observation plus a language instruction: a video generator imagines a human performing the task, a depth-aware pipeline extracts keyframes that preserve hand-object geometry, and retargeting maps them onto a 29-DoF humanoid with dexterous hands. A frozen vision-language model then acts as a calibration agent, watching real rollout traces and proposing bounded, structured edits to the keyframe skill until failures move toward task completion. After refinement the VLM leaves the loop; test-time execution re-estimates object pose and re-grounds the object-relative keyframes. If true, this matters because it replaces expensive hardware demonstration collection with generated video plus pretrained models while keeping success across object configurations and disturbances.

What carries the argument

The central object is the object-centric keyframe skill: an ordered sequence of keyframes, each specifying a stage (approach, align, fixed), a reference object, end-effector poses for both arms, dexterous-hand commands, and a validity mask. The argument is carried by three linked mechanisms: depth-aware reconstruction and retargeting that recover metric wrist and hand motion from the generated video; a frozen VLM policy editor whose outputs are restricted to structured operations (keep, align, offset, insert, delete) and validated by a deterministic projector; and test-time object-centric re-grounding, where each keyframe target is rebuilt from the current estimated object pose and a learned relative transform. The projection step maintains feasibility throughout, so the VLM reasons about what should change while deterministic geometry decides whether the proposed change may be executed.

What would settle it

Run RoboReact on the same four tasks with the VLM editor disabled or replaced by random bounded edits of the same schema, keeping compilation, re-grounding, and the whole-body controller unchanged. If terminal success on tasks such as Pour Water or Open Box stays near the reported levels, the claimed agentic refinement is not load-bearing; if it collapses, the calibration loop is responsible for the result.

Watch

Extended reading notes

Core claim

The central claim is that a non-metric generated human video has enough information to compile a deployable whole-body skill once the interaction structure is re-expressed object-centrically and calibrated physically. On four long-horizon bimanual tasks (Hand Over, Pour Water, Open Box, Open Drawer), RoboReact reaches an average terminal success rate of 81.3% over 20 trials per task, matching the 80.0% of a one-shot real-video prior and beating the ReKep and YOTO baselines, with no task-specific demonstration recordings. The paper further reports that under two-stage disturbances such as a lowered table, an object shift, and a base-pose perturbation, the frozen execution stack retains 80–94% of nominal average completed length, and that performance improves monotonically with refinement budget and with stronger video generators and VLM editors.

Load-bearing premise

The whole method stands on the assumption that a frozen vision-language model can reliably turn a handful of failed robot rollouts into small, correct fixes to the keyframe plan within twenty tries; if it cannot, the robot is left imitating a generated video it never calibrated.

Editorial extensions

If this is right

  • Whole-body bimanual skills can be bootstrapped from one RGB-D frame and a language prompt, so scaling to new tasks can follow the same compile-calibrate-freeze pipeline without collecting demonstrations for each task.
  • Performance improves as the video generator and VLM editor get stronger, so further gains should come from foundation-model progress rather than from new robot-specific data.
  • The VLM only needs to be in the loop during calibration: once the skill is frozen, it keeps 80–94% of nominal performance under disturbances, meaning test-time control can remain cheap and fast.
  • With more refinement rounds, the failure front shifts from early reachability to downstream contact, manipulation, and release phases, indicating that calibration primarily improves interaction quality rather than just reaching.
  • The one-shot real-video prior is nearly matched, supporting the design premise that generated videos supply task order and hand-object interaction structure rather than metric trajectories.
  • If the central claim holds, the practical cost of teaching a humanoid a new manipulation task drops to acquiring a single RGB-D image and writing a sentence, with the rest automated by pretrained models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication left implicit is that the VLM refinement loop, treated as an in-context optimizer, has no stated convergence guarantee; a natural extension would be to characterize whether the policy sequence is monotone or to identify failure modes where valid edits cycle.
  • The results are reported on one robot platform with a fixed whole-body controller, so the strongest cross-embodiment test would be transferring a distilled skill to a humanoid with different kinematic limits and observing whether the projection step alone can keep it feasible.
  • A testable extension is to decouple the VLM's diagnostic labels (success, alignment, grasp, contact, infeasible) from its structured edit operations, and measure which misdiagnoses most degrade the refinement loop, isolating whether the bottleneck is perception or proposal generation.
  • The disturbance results suggest that object-centric re-grounding is what confers stability once the robot is in a valid operating region; one could stress this by perturbing the object pose at every keyframe rather than only before manipulation, to find where re-anchoring alone breaks down.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. RoboReact proposes a pipeline that, from a single egocentric RGB-D frame and a language instruction, generates a human manipulation video, selects semantic keyframes with a VLM, compiles them into an object-centric whole-body keyframe skill for a humanoid, refines the skill through VLM-guided calibration rollouts, and freezes it for deployment with online pose re-grounding. Experiments on a 29-DoF Unitree G1 with two dexterous hands across four long-horizon bimanual tasks report an 81.3% mean terminal success rate with one frozen skill per task, roughly matching a one-shot real-human-video prior and exceeding ReKep and YOTO. The paper also studies the effects of refinement budget, VLM editor capability, video-generator quality, and ablation of keyframe selection, rollout memory, and third-person camera.

Significance. If the central claim holds, RoboReact is a genuinely useful step toward scalable humanoid skill acquisition: it replaces task-specific demonstration collection with generated video plus closed-loop refinement, and it demonstrates real-robot generalization across object poses, backgrounds, and disturbances. The paper's strengths include real-hardware evaluation over four long-horizon tasks, a clear ablation structure, and a scaling study showing sensitivity to VLM and video-generator capability. However, the headline autonomy claim is currently not fully supported because the calibration protocol permits up to five sparse human hints per skill-distillation run and no experiment removes them; the statistical evidence is also thin, with most comparisons resting on 13-20 trials per cell and no confidence intervals or significance tests.

major comments (3)
  1. [Sec. 4.1 / Eq. (6)] The calibration protocol explicitly permits up to five sparse human hints per skill-distillation run, and Eq. (6) passes the hint u_n into the frozen VLM editor A_ICL. The abstract and Sec. 1 claim that RoboReact works 'without requiring teleoperation or human demonstrations,' but no experiment with max_hints=0 is reported, and Table 4 ablates keyframe selection, rollout memory, and third-person camera but not the human hints. Because the one-shot real-video prior also uses the same hint protocol, the comparisons in Table 1 do not isolate the hint effect either. If these hints are necessary for the reported 81.3% success, the method is not autonomous from a single observation, and the central claim is overstated. The authors should either run a max_hints=0 ablation or substantially revise the autonomy claim.
  2. [Sec. 4.2 / Tables 1, 3, and 5] All headline performance comparisons are based on small sample sizes (13-20 trials per cell) with no confidence intervals, error bars, or significance tests. In Table 1, RoboReact's mean success rate of 81.3% is within a few trials of the 80.0% one-shot real-prior baseline, and on Hand Over, Open Box, and Open Drawer the two methods tie exactly; only Pour Water differs by one trial (17/20 vs 18/20). Table 5 reports a 92.3% vs 84.6% generator comparison on 13 trials, which binomial noise alone can easily produce. The conclusion in Sec. 5 that RoboReact 'consistently outperforms' baselines is therefore not statistically supported for the real-video prior. Please report confidence intervals or exact tests, and soften comparative claims where the evidence is indistinguishable.
  3. [Sec. 3.3 / Algorithm 1] The refinement loop is described as a 'sequential constrained skill distillation' and in-context optimization, but the paper provides no convergence analysis, no monotonic-improvement guarantee, and no study of what happens when the VLM editor proposes no useful edits within the Nmax = 20 budget. Algorithm 1 returns the current skill when an edit is invalid or projection fails, yet a valid edit can still fail to improve task success or can fix one step while breaking another. Since the deterministic validator checks schema, cited grounding, offset bounds, and keyframe structure rather than semantic progress, the reader cannot tell whether the reported gains come from the VLM's editing or from the sparse human hints. I recommend reporting per-round edit acceptance, edit type, success deltas, and stagnation cases, or at least adding a sensitivity analysis over Nmax for the stronger editor.
minor comments (6)
  1. [Sec. 4.2 (Q3)] The table references in this section appear swapped: the first sentence 'Table 5 reveals a clear hierarchy among the complementary information sources' refers to the ablation results in Table 4, while the later sentence 'Table 5 provides a complementary upstream analysis' refers to the video-generator comparison in Table 5. Please fix the cross-references.
  2. [Table captions / Sec. 4.1] The number of trials per cell varies across tables (20 in Table 1, 13 in Tables 2-6) without a consolidated explanation. State the trial count in each table caption and clarify why the refinement-budget and ablation studies use 13 trials rather than 20.
  3. [Sec. 3.2 / Eq. (3)] The same frozen VLM is used for semantic keyframe selection, for the semantic preservation objective Lsem in Eq. (3), and for policy edits in Eq. (6). Task success is scored externally by step completion, so the headline metric is not circular, but the semantic objective and the editor are not independent signals. A sentence acknowledging this coupling and explaining how the external step-level scoring constrains it would improve clarity.
  4. [Sec. 4.1] The one-shot real-human-video prior is said to be a 'task-specific recording' that replaces the generated video while retaining the same pipeline. Please clarify how this recording is obtained, whether it includes the same hint protocol, and whether it is a human video or a robot-kinesthetic demonstration, since this directly affects the interpretation of the comparison.
  5. [Sec. 4.1 / Hardware] The low-level controller is HOMIE, which is originally described as an isomorphic exoskeleton cockpit teleoperation system. Since the paper emphasizes that RoboReact requires no teleoperation, please state explicitly whether HOMIE's own training used teleoperated data and whether that is considered outside the task-specific teleoperation claim.
  6. [Sec. 4.2 / Q4] The robustness study in Table 6 is only run on Pour Water and only with 13 trials per case. Consider reporting at least one additional task or providing pooled confidence intervals to support the general 'robustly recovers from execution disturbances' claim in the abstract.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central result is externally grounded, though the same VLM serves as both semantic evaluator and policy editor, and the un-ablated human hints are a confound rather than a definitional reduction.

full rationale

RoboReact's derivation chain is self-contained against external benchmarks: the generated video comes from an off-the-shelf generator, metric reconstruction and object pose estimation are external (VGGT-Ω, WildDet3D), retargeting preserves interaction geometry, and the final skill is scored by external step completion (Tables 1 and 6), not by the VLM's own judgment. The refinement loop (Eq. 6) uses a frozen VLM as an in-context editor, and the same VLM also defines the semantic discrepancy term L_sem (Eq. 3); in principle this couples the optimizer to its own objective, but the loop's stopping criterion and reported metrics are external, and deterministic validity checks bound the edits, so no result reduces to a self-evaluation by construction. The sparse human hints permitted in Sec. 4.1 are a real confound for the 'single RGB-D observation' claim and should have been ablated with max_hints=0, but a natural-language failure description without policy-edit commands is not a fitted parameter or a demonstration, so this is an overclaim or correctness risk rather than circularity. No load-bearing self-citation or imported uniqueness theorem appears in the argument. Overall score 1 reflects the mild self-referential VLM design, not a definitional reduction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework does not introduce a new physical entity or new conservation law; it is an engineering pipeline. The parameters listed are hand-chosen knobs whose values are not given in the manuscript. The assumptions are standard domain assumptions in robot learning, but none are validated with error analysis, which matters because the entire claim depends on the reliability of upstream generative and perception models.

free parameters (6)
  • Loss weights lambda_s, lambda_g, lambda_m
    Chosen by hand in Eq. (2); values are not reported and could affect the semantic-geometric-motion trade-off.
  • Balance weight alpha in L_mot
    Hand-command versus wrist-pose weighting in Eq. (5); not specified.
  • Keyframe sampling rate and keyframe count = 10 Hz; K not reported
    The VLM selects an ordered subset from uniformly sampled frames; the number and spacing of keyframes affect skill expressiveness.
  • Refinement budget Nmax = 20 in main evaluation; 5/10/15/20 in ablations
    The number of calibration rollouts directly determines how much the VLM can correct the skill; performance is non-decreasing in budget.
  • Edit offset bounds and clearance biases
    Bounded pose/hand corrections and object-conditioned clearance biases are mentioned but their magnitudes are not given.
  • Human hint budget = at most 5
    The number of allowed human hints during calibration is a hand-set experimental parameter with unmeasured effect.
assumptions (5)
  • domain assumption The generated human video is a reliable prior for task semantics and hand-object interaction geometry, despite being non-metric and potentially physically implausible.
    The whole pipeline conditions on V(I1, Ptask) and distills interaction structure from it (Sections 3.1-3.2); Table 5 shows performance drops when the generator yields implausible hand scales.
  • domain assumption Hand-pose estimation and depth-aware 3D reconstruction recover metric wrist and hand motion from generated video frames.
    Section 3.2 uses recovered wrist/hand priors for retargeting; no accuracy evaluation on generated video is provided.
  • domain assumption Online object pose estimation (WildDet3D) is accurate enough for re-grounding object-relative keyframes at test time.
    Section 4.1 states WildDet3D provides online object poses; no pose error analysis is given and re-grounding correctness depends on it.
  • domain assumption The low-level whole-body controller HOMIE can track commanded base, body-height, torso, arm, and hand targets while maintaining balance.
    Section 3.4 relies on HOMIE for execution; the paper does not analyze tracking errors or balance margins.
  • domain assumption The frozen VLM can accurately diagnose rollouts and propose bounded, valid edits that improve the skill.
    Algorithm 1 and Eq. (6) make this the optimizer; there is no guarantee of convergence or correctness, and the paper only tests two VLM versions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RoboReact: Agentic Skill Distillation from Generated Egocentric Videos for Generalizable Whole-Body Manipulation." pith.science (2026). https://pith.science/paper/CCFLPONX

@misc{pith2026260803387,
  author       = {Pith},
  title        = {Pith review of: RoboReact: Agentic Skill Distillation from Generated Egocentric Videos for Generalizable Whole-Body Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCFLPONX}},
  note         = {Machine review of arXiv:2608.03387}
}
read the original abstract

Humanoid robots have the potential to perform dexterous manipulation in human environments, yet acquiring diverse and generalizable skills remains costly due to expensive hardware data collection and labor-intensive annotation. Recent advances in video generative models provide a promising opportunity to synthesize rich manipulation experiences from visual observations, but transferring such imagined behaviors into executable whole-body humanoid skills remains largely unexplored. In this work, we present RoboReact, a framework that automatically synthesizes whole-body humanoid manipulation skills from a single egocentric RGB-D observation. RoboReact generates human manipulation videos, extracts geometry-preserving interaction keyframes through depth-aware 3D reconstruction, and retargets them to high-DoF humanoid platforms while preserving hand-object interaction geometry. To bridge the gap between imagined plans and physical execution, RoboReact performs online object-centric re-grounding and leverages a vision-language model-guided refinement loop to adapt skills under geometric mismatch and execution deviations. The refined skills are executed through a whole-body controller, enabling coordinated whole-body manipulation and dexterous interaction. Experiments on real humanoid robots demonstrate that RoboReact generalizes across diverse object configurations and robustly recovers from execution disturbances without requiring teleoperation or human demonstrations. These results highlight the potential of combining generative models, vision-language reasoning, and closed-loop control for scalable humanoid skill acquisition.

Figures

Figures reproduced from arXiv: 2608.03387 by the authors.

Figure 1
Figure 1. RoboReact distills a generated interaction video [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of RoboReact. A VLM-selected generated video is compiled into an object-centric keyframe policy, refined from calibration rollouts, and executed through online object-pose re-grounding without test-time VLM access. 3.2 Constrained Skill Distillation Formulation We represent a bimanual robot skill as an ordered sequence of K keyframes, Π = {(ρk, ok, Tl a,k, Tr a,k, hl k , hr k , mk)} K k=1, (1) where ρk ∈ {a… view at source ↗
Figure 3
Figure 3. Representative semantic keyframe comparisons. The top row shows generated-video priors and the bottom row shows the corresponding real-robot rollout frames. Aligned keyframes are used by default before and during contact, while fixed keyframes retain robot-frame commands without a meaningful visible object anchor. Motion Prior Objective. Generated videos provide a soft prior over reaching direction, wrist orientatio… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The required RGB-D source is not tied to a fixed [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Each four-frame sequence shows contact with the lid, the fully opened box after releasing the lid, a stable object grasp, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The robot first squats down to reach a manipulation [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 13 linked inside Pith

  1. [4]

    Huang,W.;Wang,C.;Li,Y.;Zhang,R.;andFei-Fei,L.2024

    Egodex: Learning dexterous manipula- tion from large-scale egocentric video.arXiv preprint arXiv:2505.11709. Huang,W.;Wang,C.;Li,Y.;Zhang,R.;andFei-Fei,L.2024. ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation. InConference on Robot Learning (CoRL). Huang, W.; Wang, C.; Zhang, R.; Li, Y.; Wu, J.; and Fei- Fei, L

    arXiv 2024

  2. [5]

    Kareer, S.; Patel, D.; Punamiya, R.; Mathur, P.; Cheng, S.; Wang,C.;Hoffman,J.;andXu,D.2024

    WildDet3D: Scaling Promptable 3D Detection in the Wild.arXiv preprint arXiv:2604.08626. Kareer, S.; Patel, D.; Punamiya, R.; Mathur, P.; Cheng, S.; Wang,C.;Hoffman,J.;andXu,D.2024. EgoMimic:Scaling ImitationLearningviaEgocentricVideo.arXiv:2410.24221. Kerr, J.; Kim, C. M.; Wu, M.; Yi, B.; Wang, Q.; Goldberg, K.;andKanazawa,A.2024. RobotSeeRobotDo:Imitatin...

    arXiv 2024

  3. [6]

    Li, G.; Tsagkas, N.; Song, J.; Mon-Williams, R.; Vijayaku- mar, S.; Shao, K.; and Sevilla-Lara, L

    Hunyuan- video: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603. Li, G.; Tsagkas, N.; Song, J.; Mon-Williams, R.; Vijayaku- mar, S.; Shao, K.; and Sevilla-Lara, L. 2025a. Learning PreciseAffordancesfromEgocentricVideosforRoboticMa- nipulation. InProceedings of the IEEE/CVF International Conference on Computer Visio...

    arXiv

  4. [7]

    Liang,J.;Huang,W.;Xia,F.;Xu,P.;Hausman,K.;Ichter,B.; Florence,P.;andZeng,A.2023

    FRoM- W1: Towards General Humanoid Whole-Body Control with Language Instructions.arXiv preprint arXiv:2601.12799. Liang,J.;Huang,W.;Xia,F.;Xu,P.;Hausman,K.;Ichter,B.; Florence,P.;andZeng,A.2023. CodeasPolicies:Language Model Programs for Embodied Control. InIEEE Interna- tional Conference on Robotics and Automation (ICRA). Lin, T.; Sachdev, K.; Fan, L.; M...

    arXiv 2023

  5. [8]

    arXiv preprint arXiv:2505.11865

    GLOVER++: Unleashing the Potential of Affordance Learning from Human Behaviors for Robotic Manipulation. arXiv preprint arXiv:2505.11865. OpenAI

    arXiv

  6. [9]

    https://openai.com/index/gpt-5-6/

    GPT-5.6: Frontier Intelligence That Scales withYourAmbition. https://openai.com/index/gpt-5-6/. Ac- cessed: 2026-07-29. Pan, M.; Zhang, J.; Wu, T.; et al

    2026

  7. [10]

    Patel, S.; Mohan, S.; Mai, H.; Jain, U.; Lazebnik, S.; and Li, Y

    OmniManip: To- wards General Robotic Manipulation via Object-Centric In- teraction Primitives as Spatial Constraints.arXiv preprint arXiv:2501.03841. Patel, S.; Mohan, S.; Mai, H.; Jain, U.; Lazebnik, S.; and Li, Y

    arXiv

  8. [11]

    Qiu,R.-Z.;Yang,Y.;Cheng,X.;etal.2024.WildLMa:Long- Horizon Loco-Manipulation in the Wild

    Robotic Manipulation by Imitating Generated Videos Without Physical Demonstrations.arXiv preprint arXiv:2507.00990. Qiu,R.-Z.;Yang,Y.;Cheng,X.;etal.2024.WildLMa:Long- Horizon Loco-Manipulation in the Wild. InarXiv preprint arXiv:2411.15131. Singh, H.; et al

    arXiv 2024

Show all 18 references
  1. [12]

    Singh, H

    MALMM: Multi-Agent Large Lan- guage Models for Zero-Shot Robotic Manipulation.arXiv preprint arXiv:2411.17636. Singh, H. G.; Loquercio, A.; Sferrazza, C.; Wu, J.; Qi, H.; Abbeel, P.; and Malik, J

    arXiv

  2. [13]

    InInternational Conference on Robotics and Automation (ICRA)

    Hand-Object Interaction Pretraining from Videos. InInternational Conference on Robotics and Automation (ICRA). IEEE. Wan, T.; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen,D.;Yu,F.;Zhao,H.;Yang,J.;etal.2025. Wan:Open and advanced large-scale video generative models.arXi...

    2025 arXiv

  3. [14]

    Yang, L.; Huang, X.; Wu, Z.; Kanazawa, A.; Abbeel, P.; Sferrazza, C.; Liu, C

    VGGT-Ω.arXiv preprint arXiv:2605.15195. Yang, L.; Huang, X.; Wu, Z.; Kanazawa, A.; Abbeel, P.; Sferrazza, C.; Liu, C. K.; Duan, R.; and Shi, G. 2025a. OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene In- teraction.arXiv p...

    arXiv

  4. [15]

    arXiv preprint arXiv:2506.04227

    Object-centric 3D Motion Field for Robot Learning from Human Videos. arXiv preprint arXiv:2506.04227. Yuan, C.; Zhou, R.; Liu, M.; Hu, Y.; Wang, S.; Yi, L.; Wen, C.; Zhang, S.; and Gao, Y

    arXiv

  5. [16]

    Ze, Y.; Chen, Z.; Araújo, J

    MotionTrans: Human VR Data Enable Motion-Level Learning for Robotic Manipula- tion Policies.arXiv preprint arXiv:2509.17759. Ze, Y.; Chen, Z.; Araújo, J. P.; Cao, Z.-a.; Peng, X. B.; Wu, J.; and Liu, C. K

  6. [17]

    Zhang, X.; Qin, H.; Wang, F.; et al

    TWIST: Teleoperated Whole-Body Imitation System.arXiv preprint arXiv:2505.02833. Zhang, X.; Qin, H.; Wang, F.; et al

    arXiv

  7. [18]

    Zhou,H.;etal.2025

    LaMMA-P: Gen- eralizable Multi-Agent Long-Horizon Task Allocation and Planning with LM-Driven PDDL Planner.arXiv preprint arXiv:2409.20560. Zhou,H.;etal.2025. YouOnlyTeachOnce:LearningOne- Shot Bimanual Manipulation from Video Demonstrations. InRobotics: Science and Systems (RSS)

    2025 arXiv

  8. [2024]

    In8th Annual Conference on Robot Learning

    Object- Centric Dexterous Manipulation from Human Motion Data. In8th Annual Conference on Robot Learning. Du,Y.;Yang,S.;Dai,B.;Dai,H.;Nachum,O.;Tenenbaum, J.B.;Schuurmans,D.;andAbbeel,P.2023.LearningUniver- salPoliciesviaText-GuidedVideoGeneration. InAdvances in Neural Informa...

    2023

  9. [2025]

    Ben, Q.; Jia, F.; Zeng, J.; Dong, J.; Lin, D.; and Pang, J

    Residual off-policy rl for finetuning behaviorcloningpolicies.arXivpreprintarXiv:2509.19301. Ben, Q.; Jia, F.; Zeng, J.; Dong, J.; Lin, D.; and Pang, J. 2025.Homie:Humanoidloco-manipulationwithisomorphic exoskeleton cockpit.arXiv preprint arXiv:2502.13013. BenchCAD

    2025

  10. [2026]

    https://benchcad.com/

    BenchCAD: A Benchmark for Program- matic CAD. https://benchcad.com/. Accessed: 2026-07-29. Bharadhwaj, H.; Mottaghi, R.; Gupta, A.; and Tulsiani, S. 2024a. Track2Act: Predicting Point Tracks from Internet VideosenablesGeneralizableRobotManipulation. InEuro- pean Conference on ...

    2026 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.