Pith. sign in

REVIEW 3 major objections 4 minor 57 references

Teleopit: A Full-Embodiment Humanoid Teleoperation System

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Teleopit claims that a single VR headset can drive a humanoid's whole body, dexterous hands, and active camera, and that 96 demonstrations collected this way train policies that succeed at 90–95% on a real bottle-placement task.

desk verdict A solid, useful humanoid teleoperation systems paper whose central claim survives scrutiny, with the main soft spot being no direct quantitative real-robot tracking evaluation. read the letter →

arxiv 2608.01834 v1 pith:5C5ZTVKM submitted 2026-08-03 cs.RO

Bingqian Wu , Zicheng Xu , Xianghui Fan , Dayu Li , Xiangru Huang This is my paper
classification cs.RO
keywords humanoidteleoperationvirtualrealitywhole-bodycontroldexteroushandretargetingmotiontrackingimitationlearningloco-manipulationdemonstrationcollection
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

Teleopit argues that a single consumer VR headset can deliver everything a humanoid teleoperation system needs: body, hand, and head signals, all mapped in real time to whole-body motion, dexterous hands, and an active camera. The system's learned whole-body tracker uses a short temporal history and failure-focused rewind sampling to stay stable on noisy live references, and its hand retargeter transfers poses across different robot hands without per-hand tuning. The paper reports 91.7% and 100% tracking success on held-out mocap and live VR references in simulation, and real-robot demonstrations that combine locomotion, manipulation, and viewpoint control. In the key learning result, ACT and GR00T N1.7 policies trained on 96 demonstrations collected with Teleopit reach 90.0% and 95.0% real-world success on a bottle-placement task.

What carries the argument

The load-bearing machinery is a three-part pipeline. First, the learned whole-body tracker: a PPO policy that outputs 29 joint-target offsets toward a reference pose, with a 10-frame temporal history encoder and failure-aware rewind sampling that concentrates training on the transitions that make teleoperation fail; it is anchored at the torso and covers 14 body links. Second, the hand retargeter: an SLSQP optimizer minimizing a sum of three scale-free objectives—normalized finger-segment direction, one-sided fingertip-distance closure, and thumb-frame alignment—with the same weights, activation thresholds, and solver settings for every hand. Third, the asynchronous runtime: latest-only queu

What would settle it

Run the trained motion tracker on the real humanoid and measure success rate over a held-out set of live VR references using the paper's own termination criteria (anchor height error > 0.25 m, anchor orientation error > 1.0 rad, ankle/wrist height error > 0.25 m) instead of qualitative videos; if hardware success is well below the simulator's 91.7% (mocap) and 100% (PICO), the central claim fails. A second check: test the hand retargeter with the shared solver settings on a hand that shares the same semantic links but has substantially different bone lengths; if the shared settings fail to clo

Watch

Extended reading notes

Core claim

Teleopit claims that a consumer VR headset is a complete operator interface for full-embodiment humanoid teleoperation: body skeleton, hand keypoints, and head pose command a 29-DoF humanoid, dexterous hands of several morphologies, and a 2-DoF active camera in one loop. The body tracker is a single policy trained with a 10-step history encoder and failure-aware rewind sampling; it reports 91.7% success on mocap and 100% on live VR references in simulation. The hand retargeter is a shared SLSQP optimization of normalized finger directions, one-sided fingertip closure, and thumb-frame alignment, with fixed weights across six hand morphologies. From 96 successful bottle-placement demonstration

Load-bearing premise

The whole-body tracker is trained and evaluated in simulation on the target humanoid with domain randomization, and the real-robot tracking results are qualitative; if the real robot's dynamics, sensing, or the operator's VR reference distribution differ from the simulation more than the randomization covers, the 91.7%/100% simulator success rates will not hold on hardware.

Editorial extensions

If this is right

  • Humanoid demonstration collection can be done with nothing more than a consumer VR headset, removing the need for inertial suits and instrumented gloves.
  • Dexterous hands can be swapped or upgraded while keeping the same retargeting objective and solver settings, reducing integration effort.
  • Because the task policy commands the same reference interface that the operator drives, 96 demonstrations suffice for two different policy families to reach 90–95% success on a real loco-manipulation task.
  • Head-driven viewpoint control lets the operator maintain situational awareness during whole-body tasks, supporting long-horizon demonstration collection.
  • A single tracker policy covers static, running, sidestepping, turning, kneel-stand, and sit-stand transitions without motion-specific switching.

Reading between the lines

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

  • Failure-aware rewind sampling is a general curriculum idea: retrying the transitions that caused past failures could improve any reference-tracking reinforcement learning problem, beyond teleoperation.
  • The scale-free normalized-direction objective might extend to non-anthropomorphic hands or grippers, although the paper only tests six anthropomorphic designs.
  • The paper's tracking success figures come from simulation; a quantitative hardware evaluation with the same termination criteria would reveal how much of the sim-to-real gap remains.
  • The relative-root action space may be a key reason ACT and GR00T transfer from only 96 demonstrations, since it removes absolute-position variation across starting poses.
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 / 4 minor

Summary. Teleopit is presented as a VR-driven full-embodiment humanoid teleoperation system. It combines a learned whole-body motion tracker trained with PPO (history encoder, failure-aware rewind sampling) to map PICO body/head/hand signals to a Unitree G1, an optimization-based hand retargeter using normalized finger-direction, fingertip-distance, and thumb-frame objectives, and a 2-DoF head-driven active vision module. The paper reports tracker comparisons and ablations on held-out mocap and PICO references, cross-embodiment retargeting results on six dexterous hands, qualitative real-robot teleoperation with latency measurements, and downstream policy learning: ACT and GR00T N1.7 trained on 96 successful demonstrations achieve 90.0% and 95.0% success on a real bottle-placement task.

Significance. If the full set of claims holds, the paper makes a useful systems contribution: a commodity-VR interface for collecting whole-body, dexterous-hand, and viewpoint demonstrations on a real humanoid, with a hand retargeter that transfers across morphologies without hand-specific tuning. The downstream policy-learning experiment is a genuine external validation: the recorded demonstrations are consumed by two different policies and evaluated in closed-loop deployment on the robot. The component experiments for the retargeter (Tables 8-10) directly support the claimed design choices, and the use of held-out validation windows for the tracker is methodologically sound. The main weaknesses are the absence of quantitative sim-to-real tracking evidence on hardware and the small trial counts behind the headline success rates, both of which are load-bearing for the central claims.

major comments (3)
  1. [§4.1 Real-Robot Tracking / §3.2 Domain Randomization] The load-bearing claim that the learned tracker provides dynamically feasible whole-body control on the real G1 is not quantitatively supported. The simulator results (Table 6, 91.7% mocap / 100.0% PICO success) and the domain randomization of Table 3 are plausible, but the only hardware evidence is Figure 8, which is described qualitatively. No real-robot tracking error, repeated-trial success rate, or comparison against the Section 3.2 termination criteria is reported. Since every downstream result (teleoperation demos and autonomous policy deployment) assumes the tracker transfers to hardware, this unmeasured sim-to-real gap is central. Please add quantitative hardware tracking evaluation, e.g., n repeated trials with measured root/joint errors and termination rate, and a check of whether the Table 3 randomization covers the observed real-world reference distribution; otherwise the re
  2. [§4.4 Policy Learning, Table 13] The headline success rates are point estimates from only 20 trials per policy. For 18/20, a 95% binomial confidence interval spans roughly 68-99%; for 19/20, roughly 75-99.9%. The abstract and conclusion state '90.0% and 95.0%' with no uncertainty, and the comparison between ACT and GR00T N1.7 cannot be assessed from these numbers. Please report confidence intervals and/or increase the number of evaluation trials; at minimum, avoid implying a reliable ordering between the two policies. The same concern applies to the 96/100 teleoperation yield rate, which is also a single binomial estimate.
  3. [§4.1 Table 7] The ablation results are reported without repeated-seed variance. Full (reduced) attains 74.0% success versus 72.9% without rewind sampling and 73.5% without the history encoder; these are gaps of 0.5-1.1 percentage points and may be within run-to-run noise. Moreover, 'w/o rewind' improves root-position and root-velocity errors, complicating the interpretation. The claim that the history encoder and failure-aware rewind sampling improve tracking is not established by the current evidence. Please provide multiple training seeds with mean ± std (or a significance test) for the ablations, or soften the corresponding claims.
minor comments (4)
  1. [§5 Conclusion] The conclusion states that 'retargeting experiments cover more than a dozen dexterous hands,' but Table 8 evaluates six hands and the text elsewhere consistently says six. Please correct this overstatement.
  2. [§4.2, Eq. (4)] The activation a_k in the distance objective is said to 'increase smoothly' when the human fingertips approach a 4 cm contact threshold, but the functional form is not given. Please specify the activation formula so the retargeter is reproducible.
  3. [§4.3 End-to-End Latency] The latency values are estimated from manually labeled video events on four paths. Please report the number of repeated measurements per path and the uncertainty more explicitly; as written, the 0.05-0.15 s figures are single observations.
  4. [§4.1 Table 6] The text says Teleopit provides a 'favorable balance,' but HoloMotion has lower root-position error on both subsets and SONIC has lower MPJPE on mocap. Please state more precisely in which sense the balance is favorable, especially given the different training data and budgets of the baselines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evaluation is empirical and externally benchmarked; the unquantified sim-to-real gap is a generalization risk, not a circular derivation.

full rationale

Teleopit is an engineering/system paper whose claims are supported by held-out empirical evaluation rather than by a derivation chain that reduces to its inputs. The whole-body tracker is trained in the mjlab simulator with PPO on a mixture of public mocap data and its own PICO recordings, but the reported success rates are computed on validation windows explicitly disjoint from training, and all baselines (TWIST2, SONIC, HoloMotion) are re-evaluated under the same termination criteria; no success-rate value is fitted from or defined in terms of the reported result. The hand retargeter optimizes geometric objectives (normalized directions, fingertip distance, thumb-frame alignment) and is then evaluated with independent metrics and ablations; the fact that the metrics resemble the objective terms reflects the design intent, not a circular prediction, and the cross-embodiment claim is tested on six different hands with shared settings. The downstream ACT and GR00T N1.7 policies are trained on 96 recorded demonstrations and deployed on the real robot, giving external success rates (90% and 95%) that cannot be derived from the system's own parameters by construction. There is no load-bearing self-citation, no imported uniqueness theorem, and no ansatz smuggled in via the authors' prior work; the cited prior systems and simulator are independent external artifacts. The main weakness is the absence of quantitative real-robot tracker error and an unmeasured sim-to-real gap, which is a correctness/generalization concern, not circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The paper's load-bearing choices are hand-chosen training hyperparameters, reward weights, termination thresholds, and the assumed validity of simulation-to-real transfer and SLSQP retargeting. The central claims depend on these choices being reasonable, and the paper does not provide sensitivity analyses for most of them.

free parameters (8)
  • Tracker reward weights and Gaussian sigmas = weights and sigma values in Table 2
    Hand-chosen reward shaping; directly determines tracking behavior and success rate in Section 4.1.
  • Tracker termination thresholds = 0.25 m position, 1.0 rad orientation, 0.25 m ankle/wrist height
    Hand-chosen thresholds define the headline success rate metric (SR); changing them changes all reported numbers.
  • Rewind probability and rewind offset = rewind_prob = 0.8; random offset
    Hyperparameter of failure-aware rewind sampling; chosen by hand, only one value tested.
  • History length H = 10
    Hand-chosen window for temporal context; ablations only test with and without, not different lengths.
  • Action scale s and default pose q_def = per-joint scale and default pose for Unitree G1
    Used in Eq. (1) to decode policy actions into joint targets; chosen by hand.
  • Hand retargeter weights, activation threshold, smoothing, solver settings = shared constants: w_i, w_k, w_p, w_s, 4 cm threshold, lambda, SLSQP settings
    Central to the no-tuning claim; globally set but not learned and not compared to alternative parameter choices.
  • Domain randomization ranges = Table 3 values
    Hand-chosen ranges assumed to bridge sim-to-real; no sensitivity analysis is reported.
  • Imitation hyperparameters = batch sizes, learning rates, action chunk lengths in Table 11
    Standard training choices; not part of the central claim but needed to reproduce the 90/95 percent results.
assumptions (6)
  • domain assumption SLSQP with analytic gradients converges to a good retargeting solution for all six hands
    Hand retargeting results in Table 8 rely on this; no global optimality guarantee is given.
  • domain assumption PICO headset hand, body, and head tracking is sufficiently accurate for teleoperation
    The interface consumes 24-joint body, 26 keypoints per hand, and head pose at 60 Hz; tracking error is not quantified.
  • domain assumption mjlab simulation with Table 3 domain randomization transfers to the real Unitree G1
    Real-robot tracking section provides qualitative videos only; quantitative sim-to-real transfer is not measured.
  • standard math Gram-Schmidt orthonormalization of thumb-base axes is well-defined and meaningful across hands
    Used to construct the thumb frame in Eq. (5).
  • domain assumption PD controllers at 200 Hz track the 50 Hz policy joint targets
    Control stack assumes decimation 4 and stable tracking; no hardware PD tracking error is reported.
  • domain assumption The 52D reference action space is sufficient for the bottle-placement task and preserves demonstration information
    Policy learning in Section 4.4 depends on this; no analysis of lost information is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teleopit: A Full-Embodiment Humanoid Teleoperation System." pith.science (2026). https://pith.science/paper/5C5ZTVKM

@misc{pith2026260801834,
  author       = {Pith},
  title        = {Pith review of: Teleopit: A Full-Embodiment Humanoid Teleoperation System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5C5ZTVKM}},
  note         = {Machine review of arXiv:2608.01834}
}
read the original abstract

Humanoid teleoperation for demonstration collection requires coordinated whole-body motion, continuous dexterous hand control, and viewpoint control. Existing systems either simplify hand commands or depend on dedicated wearable sensors for fine-grained hand motion. We introduce Teleopit, a full-embodiment teleoperation system that maps body, hand, and head signals from VR to a humanoid body, configurable dexterous hands, and a 2-DoF active vision module. A history encoder and failure-aware rewind sampling improve the motion tracker on both motion-capture and live VR references. An optimization-based hand retargeter combines normalized finger directions, fingertip closure, and thumb-frame alignment to map human hand motion to different dexterous hands without tuning hand-specific objective or solver hyperparameters. Component experiments evaluate tracking success rate and retargeting behavior, while real-robot teleoperation demonstrates coordinated locomotion, manipulation, and viewpoint control. ACT and GR00T N1.7 policies trained on 96 successful demonstrations collected with Teleopit achieve task success rates of 90.0% and 95.0%, respectively, when deployed on the humanoid. The project page is available at https://botrunner64.github.io/teleopit-page.

Figures

Figures reproduced from arXiv: 2608.01834 by the authors.

Figure 1
Figure 1. Teleopit uses VR body, hand, and head tracking for full-embodiment humanoid teleoperation. The interface coordinates dynamically feasible whole-body control, configurable dexterous hands, and active vision during real-world loco-manipulation. Abstract Humanoid teleoperation for demonstration collection requires coordinated whole-body motion, continuous dexterous hand control, and viewpoint control. Existing systems … view at source ↗
Figure 2
Figure 2. Whole-body motion tracking pipeline. The actor combines current proprioception and temporal history, while the training-only critic also uses privileged state. PPO opti￾mizes predicted joint targets. New hands require semantic link mapping but no hand-specific objective or solver hyperparameter tuning. 3 Method 3.1 Overview Teleopit maps body, hand, and head signals from PICO to a humanoid robot in real time. The in… view at source ↗
Figure 4
Figure 4. Hand retargeting objectives. Direction, fingertip￾distance, and thumb-frame objectives map the human pose to a robot hand with different morphology. Direction Objective. The orange arrows in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (11 more)
Figure 7
Figure 7. Figure 7: Active vision module. CAD and robot-mounted 2-DoF yaw–pitch camera. native rates. PICO signals, hand retargeting, and viewpoint control run at 60 Hz; the body policy runs at 50 Hz; low-level body control runs at 200 Hz; and recording runs at 30 Hz. Latest-only communic…
Figure 5
Figure 5. Figure 5: Deployment architecture. Asynchronous processes map PICO body, hand, and head streams to robot commands while returning video and recording synchronized data [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: PICO motion interface. Host-side visualization of the body skeleton, head pose, and articulated hand keypoints reconstructed from the headset and ankle-tracker signals. freshness checks described below, the sensing, control, video, and recording processes run asynchron…
Figure 8
Figure 8. Figure 8: Real-robot whole-body tracking. Static motions and continuous kneel–stand and sit–stand transitions on hardware [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Pose fidelity across gestures. Human skeletons and retargeted poses show finger-direction, pinch, and thumb￾opposition fidelity [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 11
Figure 11. Figure 11: Retargeting ablation. The distance objective maintains pinch closure, and the frame objective preserves thumb opposition. Distance objective [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Video-based estimates of end-to-end latency. Paired images show the initiating event (top) and the downstream arrival (bottom). The estimated delays are approximately 0.10 s for whole-body response, 0.05 s for viewpoint-control response, 0.10 s for the video stream ar…
Figure 13
Figure 13. Figure 13: Head-driven viewpoint control. Views returned to the headset as the operator looks around from the neutral view (center). The eight surrounding frames show the camera coverage obtained by commanding yaw and pitch through head motion. Observation and recorded reference…
Figure 14
Figure 14. Figure 14: Integrated loco-manipulation. Body, hand, and vision control across four household tasks. Imitation Policy General Motion Tracker Root Pos. Root Quat. Body Joint Pos. Hand Joint Pos. Neck Joint Pos. Rel. Root X/Y Pos. Abs. Root Height Rel. Root 6D Rot. Body Joint Pos.…
Figure 15
Figure 15. Figure 15: Hierarchical interface for autonomous execution. The task policy predicts a 52D reference motion chunk from the robot image and measured joint state. The initial root pose converts the relative root prediction back to a 50D robot reference. Root and body references en…
Figure 16
Figure 16. Figure 16: Teleoperation collection and autonomous deployment. Each row proceeds from left to right. (a) Teleopit records an operator demonstration of grasping a bottle and placing it inside the adjacent box. (b) GR00T N1.7 autonomously executes the same task. (c,d) Without furt…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 4 canonical work pages

  1. [1]

    Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    Anthony Brohan, Noah Brown, Justice Carbajal, Yev- gen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    arXiv 2022

  2. [2]

    Rt-2: Vision-language- action models transfer web knowledge to robotic control

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. Rt-2: Vision-language- action models transfer web knowledge to robotic control. InConference on Robot Learning, pages 2165–2183. PMLR, 2023

    2023

  3. [3]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0

    Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Ab- hishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

    2024

  4. [4]

    Droid: A large-scale in-the-wild robot manipulation dataset

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ash- win Balakrishna, Sudeep Dasari, Siddharth Karam- cheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024

    arXiv 2024

  5. [5]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, EthanFoster,GraceLam,PannagSanketi,etal. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    arXiv 2024

  6. [6]

    𝜋0: Avision- language-action flow model for general robot control

    Kevin Black, Noah Brown, Danny Driess, Adnan Es- mail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom,KarolHausman,BrianIchter,etal. 𝜋0: Avision- language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024

    arXiv 2024

  7. [7]

    GR00T n1: An open foundation model for generalist humanoid robots

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. GR00T n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025

    arXiv 2025

  8. [8]

    Open-television: Teleoperation with immersive active visual feedback.arXiv preprint arXiv:2407.01512, 2024

    Xuxin Cheng, Jialong Li, Shiqi Yang, Ge Yang, and Xiaolong Wang. Open-television: Teleoperation with immersive active visual feedback.arXiv preprint arXiv:2407.01512, 2024

    arXiv 2024

Show all 57 references
  1. [9]

    Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832, 2025

    Yanjie Ze, Siheng Zhao, Weizhuo Wang, Angjoo Kanazawa, Rocky Duan, Pieter Abbeel, Guanya Shi, Jiajun Wu, and C Karen Liu. Twist2: Scalable, portable, and holistic humanoid data collection system.arXiv preprint arXiv:2511.02832, 2025

    2025

  2. [10]

    Humdex: Humanoid dexterous manipulation made easy.arXiv preprint arXiv:2603.12260, 2026

    Liang Heng, Yihe Tang, Jiajun Xu, Henghui Bao, Di Huang, and Yue Wang. Humdex: Humanoid dexterous manipulation made easy.arXiv preprint arXiv:2603.12260, 2026

    2026

  3. [11]

    Xrobotoolkit: A cross-platform framework for robot tele- operation

    Zhigen Zhao, Liuchuan Yu, Ke Jing, and Ning Yang. Xrobotoolkit: A cross-platform framework for robot tele- operation. In2026 IEEE/SICE International Symposium on System Integration (SII), pages 15–20. IEEE, 2026

    2026

  4. [12]

    Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system

    Ankur Handa, Karl Van Wyk, Wei Yang, Jacky Liang, Yu-Wei Chao, Qian Wan, Stan Birchfield, Nathan Ratliff, and Dieter Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 91...

    2020

  5. [13]

    Anyteleop: A general vision-based dexterous robot arm- hand teleoperation system

    Yuzhe Qin, Wei Yang, Binghao Huang, Karl Van Wyk, Hao Su, Xiaolong Wang, Yu-Wei Chao, and Dieter Fox. Anyteleop: A general vision-based dexterous robot arm- hand teleoperation system. InProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July

  6. [14]

    Geometricretargeting: Aprincipled,ultrafast neural hand retargeting algorithm

    Zhao-Heng Yin, Changhao Wang, Luis Pineda, Krishna Bodduluri, Tingfan Wu, Pieter Abbeel, and Mustafa Mukadam. Geometricretargeting: Aprincipled,ultrafast neural hand retargeting algorithm. In2025 IEEE/RSJ In- ternationalConferenceonIntelligentRobotsandSystems (IROS), pages 173...

    2025

  7. [15]

    mjlab: A lightweightframeworkforgpu-acceleratedrobotlearning

    Kevin Zakka, Qiayuan Liao, Brent Yi, Louis Le Lay, Koushil Sreenath, and Pieter Abbeel. mjlab: A lightweightframeworkforgpu-acceleratedrobotlearning. arXiv preprint arXiv:2601.22074, 2026

    2026

  8. [16]

    Nathan Miller, Odest Chadwicke Jenkins, Marcelo Kall- mann, and Maja J. Mataric. Motion capture from inertial sensing for untethered humanoid teleoperation. InPro- ceedings of the IEEE-RAS International Conference on Humanoid Robots, 2004

    2004

  9. [17]

    Sreeni- vasa, and Jean-Paul Laumond

    Francisco-Javier Montecillo-Puente, Manish N. Sreeni- vasa, and Jean-Paul Laumond. On real-time whole- body human to humanoid motion transfer. InPro- ceedings of the 7th International Conference on In- formatics in Control, Automation and Robotics, pages 22–31, 2010. doi: 10.5...

    2010 doi

  10. [18]

    Learning human- to-humanoidreal-timewhole-bodyteleoperation.In2024 IEEE/RSJInternationalConferenceonIntelligentRobots and Systems (IROS), 2024

    TairanHe,ZhengyiLuo,WenliXiao,ChongZhang,Kris Kitani, Changliu Liu, and Guanya Shi. Learning human- to-humanoidreal-timewhole-bodyteleoperation.In2024 IEEE/RSJInternationalConferenceonIntelligentRobots and Systems (IROS), 2024. doi: 10.1109/IROS58592. 2024.10801984. URL https:...

    2024

  11. [19]

    Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning

    Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human- to-humanoid whole-body teleoperation and learning. In Proceedings of the 8th Conference on Robot Learning, volume 270 of...

    2025

  12. [20]

    Humanplus: Humanoid shadowing and imitation from humans

    Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. InProceedings of the 8th Conference on Robot Learning, volume 270 ofPro- ceedings of Machine Learning Research, pages 2828– 2844.PMLR,2025.URL https://...

    2025

  13. [21]

    Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833, 2025

    Yanjie Ze, Zixuan Chen, JoÃG, o Pedro AraÚjo, Zi- ang Cao, Xue Bin Peng, Jiajun Wu, and C Karen Liu. Twist: Teleoperated whole-body imitation system.arXiv preprint arXiv:2505.02833, 2025

    2025 arXiv

  14. [22]

    Deep imi- tation learning for humanoid loco-manipulation through human teleoperation

    MingyoSeo,SteveHan,KyutaeSim,SeungHyeonBang, Carlos Gonzalez, Luis Sentis, and Yuke Zhu. Deep imi- tation learning for humanoid loco-manipulation through human teleoperation. In2023 IEEE-RAS 22nd Interna- tional Conference on Humanoid Robots (Humanoids), pages 1–8, 2023. URL h...

    2023 arXiv

  15. [23]

    Homie: Humanoid loco- manipulationwithisomorphicexoskeletoncockpit.arXiv preprint arXiv:2502.13013, 2025

    Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. Homie: Humanoid loco- manipulationwithisomorphicexoskeletoncockpit.arXiv preprint arXiv:2502.13013, 2025

    2025 arXiv

  16. [24]

    Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks

    Yixuan Li, Yutang Lin, Jieming Cui, Tengyu Liu, Wei Liang, Yixin Zhu, and Siyuan Huang. Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks. In9th Annual Conference on Robot Learning, 2025

    2025

  17. [25]

    Humanoid ma- nipulation interface: Humanoid whole-body manipu- lation from robot-free demonstrations.arXiv preprint arXiv:2602.06643, 2026

    Ruiqian Nai, Boyuan Zheng, Junming Zhao, Haodong Zhu, Sicong Dai, Zunhao Chen, Yihang Hu, Yingdong Hu, Tong Zhang, Chuan Wen, et al. Humanoid ma- nipulation interface: Humanoid whole-body manipu- lation from robot-free demonstrations.arXiv preprint arXiv:2602.06643, 2026

    2026

  18. [26]

    Bifrostumi: Bridging robot-free demonstrations and humanoid whole-body manipulation.arXiv preprint arXiv:2605.03452, 2026

    ChenhaoYu,HongwuWang,YouhaoHu,JiachenZhang, Yuanyuan Li, and Shaqi Luo. Bifrostumi: Bridging robot-free demonstrations and humanoid whole-body manipulation.arXiv preprint arXiv:2605.03452, 2026

    2026 arXiv

  19. [27]

    Deepmimic: Example-guided deep reinforcement learning of physics-based charac- ter skills.ACM Transactions on Graphics, 37(4): 1–14, 2018

    Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based charac- ter skills.ACM Transactions on Graphics, 37(4): 1–14, 2018. doi: 10.1145/3197517.3201311. URL https://doi.org/10.1145/3197517.3201311

    2018

  20. [28]

    Amp: Adversarial motion pri- ors for stylized physics-based character control.ACM Transactions on Graphics, 40(4):1–20, 2021

    Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion pri- ors for stylized physics-based character control.ACM Transactions on Graphics, 40(4):1–20, 2021. doi: 10.1145/3450626.3459670. URL https://doi.org/ 10.1145/3450626.3459670

    2021

  21. [29]

    Perpetualhumanoidcontrolforreal-timesimulated avatars

    Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, etal. Perpetualhumanoidcontrolforreal-timesimulated avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895–10904, 2023

    2023

  22. [30]

    Univer- sal humanoid motion representations for physics-based control

    Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Win- kler, Jing Huang, Kris Kitani, and Weipeng Xu. Univer- sal humanoid motion representations for physics-based control. InInternational Conference on Learning Repre- sentations, volume 2024, pages 56766–56782, 2024

    2024

  23. [31]

    Expressive whole- body control for humanoid robots.arXiv preprint arXiv:2402.16796, 2024

    Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, and Xiaolong Wang. Expressive whole- body control for humanoid robots.arXiv preprint arXiv:2402.16796, 2024

    2024 arXiv

  24. [32]

    Ex- body2: Advanced expressive humanoid whole-body control.arXiv preprint arXiv:2412.13196, 2024

    Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Ex- body2: Advanced expressive humanoid whole-body control.arXiv preprint arXiv:2412.13196, 2024

    2024 arXiv

  25. [33]

    Hover: Versatile neural whole- body controller for humanoid robots

    Tairan He, Wenli Xiao, Toru Lin, Zhengyi Luo, Zhenjia Xu, Zhenyu Jiang, Jan Kautz, Changliu Liu, Guanya Shi, Xiaolong Wang, et al. Hover: Versatile neural whole- body controller for humanoid robots. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages ...

    2025

  26. [34]

    Gmt: General motiontrackingforhumanoidwhole-bodycontrol.arXiv preprint arXiv:2506.14770, 2025

    Zixuan Chen, Mazeyu Ji, Xuxin Cheng, Xuanbin Peng, Xue Bin Peng, and Xiaolong Wang. Gmt: General motiontrackingforhumanoidwhole-bodycontrol.arXiv preprint arXiv:2506.14770, 2025

    2025 arXiv

  27. [35]

    Track any motions un- der any disturbances.arXiv preprint arXiv:2509.13833, 2025

    Zhikai Zhang, Jun Guo, Chao Chen, Jilong Wang, Chenghuai Lin, Yunrui Lian, Han Xue, Zhenrong Wang, Maoqi Liu, Jiangran Lyu, et al. Track any motions un- der any disturbances.arXiv preprint arXiv:2509.13833, 2025

    2025

  28. [36]

    Robustandgeneralizedhumanoid motion tracking.arXiv preprint arXiv:2601.23080, 2026

    YubiaoMa,HanYu,JiayinXie,ChangtaiLv,QiangLuo, Chi Zhang, Yunpeng Yin, Boyang Xing, Xuemei Ren, andDongdongZheng. Robustandgeneralizedhumanoid motion tracking.arXiv preprint arXiv:2601.23080, 2026

    2026

  29. [37]

    Telegate: Whole-body humanoid teleoperation via gated expert selection with motion prior.arXiv preprint arXiv:2602.09628, 2026

    Jie Li, Bing Tang, and Feng Wu. Telegate: Whole-body humanoid teleoperation via gated expert selection with motion prior.arXiv preprint arXiv:2602.09628, 2026

    2026 arXiv

  30. [38]

    Sonic: Supersizing motiontrackingfornaturalhumanoidwhole-bodycontrol

    Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Fernando Castañeda, Sirui Chen, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, et al. Sonic: Supersizing motiontrackingfornaturalhumanoidwhole-bodycontrol. arXiv preprint arXiv:2511.07820, 2025. 16

    2025 arXiv

  31. [39]

    Holomotion-1 techni- cal report.arXiv preprint arXiv:2605.15336, 2026

    Maiyue Chen, Kaihui Wang, Bo Zhang, Xihan Ma, Zhiyuan Yang, Yi Ren, Qijun Huang, Zihao Zhu, Yucheng Wang, and Zhizhong Su. Holomotion-1 techni- cal report.arXiv preprint arXiv:2605.15336, 2026

    2026 arXiv

  32. [40]

    Calibrating human hand for teleoperating the hit/dlr hand

    Haiying Hu, Xiaohui Gao, Jiawei Li, Jie Wang, and Hong Liu. Calibrating human hand for teleoperating the hit/dlr hand. InProceedings of the IEEE Interna- tionalConferenceonRoboticsandAutomation,volume5, pages 4571–4576, 2004. doi: 10.1109/ROBOT.2004. 1302438. URL https://doi.o...

    2004 arXiv

  33. [41]

    Intuitive hand teleoperation by novice operators using a continuous teleoperation subspace

    Cassie Meeker, Thomas Rasmussen, and Matei Ciocar- lie. Intuitive hand teleoperation by novice operators using a continuous teleoperation subspace. In2018 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 5821–5827, 2018. doi: 10. 1109/ICRA.2018.8460506....

    2018

  34. [42]

    Combined joint- cartesian mapping for simultaneous shape and preci- sion teleoperation of anthropomorphic robotic hands

    Roberto Meattini, Dario Chiaravalli, Luigi Biagiotti, Gianluca Palli, and Claudio Melchiorri. Combined joint- cartesian mapping for simultaneous shape and preci- sion teleoperation of anthropomorphic robotic hands. IFAC-PapersOnLine, 53(2):10052–10057, 2020. doi: 10.1016/j.ifa...

    2020 doi

  35. [43]

    Human to robot hand motion mapping methods: Review and classification.IEEE Transactions on Robotics, 39(2):842–861, 2023

    Roberto Meattini, Raúl Suárez, Gianluca Palli, and Claudio Melchiorri. Human to robot hand motion mapping methods: Review and classification.IEEE Transactions on Robotics, 39(2):842–861, 2023. doi: 10.1109/TRO.2022.3205510. URLhttps://doi.org/ 10.1109/TRO.2022.3205510

    2023

  36. [44]

    Robotic telekinesis: Learning a robotic hand imitator by watching humans on youtube

    Aravind Sivakumar, Kenneth Shaw, and Deepak Pathak. Robotic telekinesis: Learning a robotic hand imitator by watching humans on youtube. InProceedings of Robotics: Science and Systems, 2022. URL https: //arxiv.org/abs/2202.10448

    2022 arXiv

  37. [45]

    Anydexrt: Calibration-free dexterous hand retargeting with few- shot human guidance.arXiv preprint arXiv:2607.08341, 2026

    Chenxi Wang, Ying Feng, Hongjie Fang, Shangning Xia, Lixin Yang, Chuan Wen, and Cewu Lu. Anydexrt: Calibration-free dexterous hand retargeting with few- shot human guidance.arXiv preprint arXiv:2607.08341, 2026

    2026 arXiv

  38. [46]

    Dexmachina: Func- tional retargeting for bimanual dexterous manipulation

    Zhao Mandi, Yifan Hou, Dieter Fox, Yashraj Narang, Ajay Mandlekar, and Shuran Song. Dexmachina: Func- tional retargeting for bimanual dexterous manipulation. arXiv preprint arXiv:2505.24853, 2025

    2025 arXiv

  39. [47]

    Spider: Scalable physics-informed dexterous retargeting.arXiv preprint arXiv:2511.09484, 2025

    Chaoyi Pan, Changhao Wang, Haozhi Qi, Zixi Liu, Homanga Bharadhwaj, Akash Sharma, Tingfan Wu, GuanyaShi,JitendraMalik,andFrancoisHogan. Spider: Scalable physics-informed dexterous retargeting.arXiv preprint arXiv:2511.09484, 2025

    2025

  40. [48]

    Dex- twist: Dexterous hand retargeting for twist motion via mixedreality-basedteleoperation

    Dongmyoung Lee, Chengxi Li, and Dongheui Lee. Dex- twist: Dexterous hand retargeting for twist motion via mixedreality-basedteleoperation. In2026IEEEInterna- tional Conference on Advanced Robotics and its Social Impacts (ARSO), pages 149–154. IEEE, 2026

    2026

  41. [49]

    Genhand: Generalised human grasp kinematic retargeting.npj Robotics, 4, 2026

    Liyuan Qi, Olaoluwa Popoola, Muhammad Ali Imran, and Wasim Ahmad. Genhand: Generalised human grasp kinematic retargeting.npj Robotics, 4, 2026. doi: 10. 1038/s44182-026-00076-1. URL https://doi.org/ 10.1038/s44182-026-00076-1

    2026 doi

  42. [50]

    Proximalpolicyoptimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford,andOlegKlimov. Proximalpolicyoptimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  43. [51]

    Dexterousteleoperationof20-dof bytedexter hand via human motion retargeting.arXiv preprint arXiv:2507.03227, 2025

    Ruoshi Wen, Jiajun Zhang, Guangzeng Chen, Zhongren Cui,MinDu, YangGou,ZhigangHan, JunkaiHu,Liqun Huang,HaoNiu,etal. Dexterousteleoperationof20-dof bytedexter hand via human motion retargeting.arXiv preprint arXiv:2507.03227, 2025

    2025 arXiv

  44. [52]

    BONES-SEED: Skeletal every- day embodiment dataset

    BONES Studio. BONES-SEED: Skeletal every- day embodiment dataset. https://bones.studio/ datasets/seed, 2026. Accessed 2026-07-14

    2026

  45. [53]

    Harvey, Mike Yurick, Derek Nowrouzezahrai, andChristopherPal

    Félix G. Harvey, Mike Yurick, Derek Nowrouzezahrai, andChristopherPal. Robustmotionin-betweening.ACM Transactions on Graphics, 39(4), 2020. doi: 10.1145/ 3386569.3392480. URL https://arxiv.org/abs/ 2102.04942

    2020

  46. [54]

    Learning fine-grained bimanual manipulation with low-cost hardware

    TonyZ.Zhao,VikashKumar,SergeyLevine,andChelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023. doi: 10.15607/RSS.2023.XIX.016

    2023 doi

  47. [55]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. InProceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition, pages 5745–5753, 2019

    2019

  48. [56]

    LeRobot: An open-source library for end- to-end robot learning

    Rémi Cadene, Simon Alibert, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, Mustafa Shukor, Jess Moss, Alexander Soare, Dana Aubakirova, Quentin Lhoest, Quentin Gallouédec, and Thomas Wolf. LeRo...

    2026

  49. [2023]

    URLhttps: //doi.org/10.15607/RSS.2023.XIX.015

    doi: 10.15607/RSS.2023.XIX.015. URLhttps: //doi.org/10.15607/RSS.2023.XIX.015

    2023 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.