Pith. sign in

REVIEW 2 major objections 4 minor 80 references

One 27-actuator body works as both a dexterous hand and a desktop humanoid, with grasping, locomotion, and a long task that switches between them.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 21:04 UTC pith:QLWE37CX

load-bearing objection A genuinely new reconfigurable hand/humanoid platform with solid real-world grasping data, but the 'RL locomotion' validation is only simulated; real-world walking is open-loop keyframes. the 2 major comments →

arxiv 2607.16187 v1 pith:QLWE37CX submitted 2026-07-17 cs.RO

Handroid: Bridging Dexterous Hand and Humanoid

Ruogu Li , Chenyang Ma , Sikai Li , Zhenyu Wei , Yunchao Yao , Haochen Shi , C. Karen Liu , Shuran Song
show 1 more author
Mingyu Ding
This is my paper
classification cs.RO
keywords Handroiddexterous handhumanoid robotmorphology reconfigurationmodule reuserobot learningloco-manipulationcross-embodiment learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Handroid is a desk-sized robot (0.33 m, 2.05 kg) whose 27 actuated joints can be rearranged into a five-finger dexterous hand (20 joints) or a small humanoid with head, arms, and legs (25 joints plus two sliding reconfiguration joints). The paper's central thesis is that a single electromechanical body, sensing backbone, and control stack can support both contact-rich manipulation and whole-body mobility, rather than building hands and humanoids as separate machines. In experiments the hand embodiment reaches 72% average success on real-world grasping across 10 objects, performs VR-teleoperated tasks and simulated-to-real in-hand cube reorientation, while the humanoid embodiment walks with simulated reinforcement-learning policies and executes authored keyframe motions on the physical robot. A long-horizon demo chains embodiment switching, walking, box pushing, docking with a robot arm, and dexterous pick-and-place. If correct, this offers a compact, reproducible way to study morphology reuse and cross-embodiment learning.

Core claim

Handroid's central claim is that one 27-actuator electromechanical body can be reused across two embodiments without hardware replacement. In the hand configuration, 20 actuated joints form an anthropomorphic five-finger structure approximating a 21-joint human-hand model; in the humanoid configuration, the same articulated module groups form a 4-joint head, two 4-joint arms, two 6-joint legs, and a 1-joint hip, with two prismatic sliding joints translating modules between layouts. The paper shows that a unified control and learning stack—VR teleoperation, object-conditioned imitation learning, reinforcement-learned in-hand reorientation, planner-guided and reference-free locomotion policies

What carries the argument

The load-bearing mechanism is module reuse: five articulated module groups serve as fingers in the hand and as head, arms, and legs in the humanoid, mounted on a central base and hip, with two rack-and-pinion prismatic joints that physically slide modules between configurations. This 'palm-to-finger is like torso-to-limb' analogy, plus a shared control stack, is what lets one body and one interface serve both manipulation and locomotion.

Load-bearing premise

The claim that the humanoid can walk rests on the assumption that simulated reinforcement-learning locomotion plus real-world open-loop keyframe playback is adequate evidence of real-world walking; if the simulator overestimates stability or the keyframe gaits are too fragile, the locomotion half of the central claim collapses even though the manipulation results stand.

What would settle it

Take the learned velocity-control policy that tracks 0.20 m/s in simulation and run it on the physical humanoid embodiment: if it cannot maintain balance and follow the commanded velocity over several steps, the claim of RL-based real-world locomotion is falsified. Alternatively, run the long-horizon embodiment-switch task many times and count failures during reconfiguration or docking to test reliability of those mechanisms.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Morphology reuse can be realized in working hardware: a dexterous hand and a humanoid from one electromechanical body, with no part replacement.
  • A common control and learning stack lets manipulation and locomotion policies share sensing, actuation, and deployment interfaces, simplifying cross-embodiment learning.
  • The long-horizon demo shows embodiment switching can be integrated into a single task workflow alongside locomotion, docking, and manipulation.
  • Simulation-trained locomotion policies (tracking and velocity control) suggest the modular body can support learning-based gait generation, not just scripted playback.
  • The 72% average grasping success across varied objects suggests a reconfigurable hand retains practical dexterity comparable to dedicated dexterous hands.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The paper's strongest untested step is sim-to-real transfer of the learned locomotion policies; deploying the velocity policy on the physical humanoid and achieving stable steps would make Handroid a low-cost testbed for cross-embodiment policy transfer, but the paper leaves that deployment undone.
  • Editorial inference: Because the same modules act as fingertips and feet and carry inertial sensors, a natural extension is sharing contact-estimation signals between manipulation and locomotion policies; the paper does not explore this.
  • Editorial inference: The reconfiguration mechanism's long-term repeatability (positional drift after many switches) is not measured; if drift stays low, the design principle could extend to other module-based morphologies beyond hand and humanoid.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper introduces Handroid, a 27-DoF desktop-scale robot whose electromechanical modules can be reconfigured into a dexterous hand (20 DoFs) or a humanoid (25 articulated DoFs plus two prismatic joints). The authors describe the mechanical/electrical design, a shared control stack, and experiments covering VR teleoperation, object-conditioned diffusion grasping (72% average success over 10 objects), real-world in-hand cube reorientation, MuJoCo-trained RL tracking and velocity-control locomotion, keyframe-authored real-world humanoid motions, and a long-horizon task that combines embodiment switching, locomotion, docking, and pick-and-place. The central claim is that a single reconfigurable body supports both contact-rich manipulation and whole-body mobility.

Significance. If the results hold, Handroid is a notable open-source platform: the dual-embodiment design is genuinely novel relative to fixed-morphology hands and humanoids, and the real long-horizon demo (switch, detach, walk, push, dock, grasp, place) is nontrivial. The grasping evaluation with 10 objects and 100 demonstrations is a strength, as is the explicit reporting of simulated tracking/velocity errors. However, the paper's abstract and contributions claim validation through 'reinforcement-learning-based locomotion,' while the learned locomotion policies are evaluated only in MuJoCo; physical humanoid motions are open-loop keyframe playback. This mismatch is the main technical concern. The paper explicitly labels its simulation-only results, so the issue is one of claim calibration rather than internal inconsistency.

major comments (2)
  1. [Abstract; §5.1 (Simulated RL Tracking / Velocity Control / Real-World Keyframe Motions); §5.2] The abstract states the platform is validated through 'reinforcement-learning-based locomotion,' and the first contribution says 'validate Handroid through ... humanoid locomotion.' Yet §5.1 reports 'Simulated RL Tracking' and 'Simulated RL Velocity Control' only in MuJoCo, while 'Real-World Keyframe Motions' are explicitly direct joint-position playback that 'do not invoke either learned policy.' The long-horizon task (§5.2) does not specify whether the humanoid walking/turning/box-pushing uses a learned policy or keyframes. Consequently, the physical humanoid has not been demonstrated to walk under a closed-loop learned controller. Please either deploy the tracking or velocity policy on the real robot and report quantitative tracking/velocity errors, or revise the abstract, contributions, and conclusion to state that learned locomotion is validated in simulation and that real-world loc
  2. [§5.1, In-hand Reorientation] The in-hand cube reorientation is described only qualitatively ('We qualitatively evaluate whether the policy can maintain the cube in-hand'). Since the introduction lists 'real-world in-hand cube reorientation' among the validated capabilities, a quantitative measure (e.g., success rate over N reorientations, or angular error over time) is needed to support that claim. If the authors intend this as a qualitative demonstration, the wording should be softened.
minor comments (4)
  1. [§5.1 heading] The heading '5.1 Dexterous Hand' contains paragraphs on Simulated RL Tracking, Simulated RL Velocity Control, and Real-World Keyframe Motions, which are humanoid experiments. Rename or restructure to avoid confusion.
  2. [Eqs. (1)-(2)] The reward weights and kernel widths are listed in Appendix III, not in the main text; consider a table or explicit values near the equations.
  3. [Table 1] An average success rate line would help; currently readers compute it by hand. Also report trial count per object (apparently 10 each) in the caption.
  4. [§5.1 heading] Typo: 'Simulated RL V elocity Control' should be 'Simulated RL Velocity Control'.

Circularity Check

0 steps flagged

No significant circularity: results are empirical hardware demonstrations; RL locomotion is simulation-only but honestly labeled, not circular.

full rationale

The paper does not present a derivation whose output is equivalent to its input. The control objectives (Eqs. 1 and 2) are standard tracking and velocity-conditioned rewards with hand-set weights and kernel widths; no parameter is fitted to reproduce the reported success rates or tracking errors. The dexterous-grasping policy is trained on 100 teleoperated demonstrations and then evaluated on randomized object poses via real-world grasping trials, providing an external success-rate measurement. The in-hand reorientation policy is trained in IsaacLab and deployed on hardware, again without calibrating any learned parameter to the reported outcome. Humanoid locomotion is explicitly split: 'Simulated RL Tracking' and 'Simulated RL Velocity Control' are MuJoCo-only, while 'Real-World Keyframe Motions' are direct joint-position playback. This is an evidence gap between the abstract's stronger phrasing and the physical results, but it is not circularity: the simulation results are not claimed to be the real-world results. Self-citations in related work (e.g., ToddlerBot [41], Current-as-Touch [61], Coordex [68]) are contextual and are not used to justify Handroid's central platform claim. The hardware design, reconfiguration mechanism, and long-horizon task are self-contained empirical contributions. Therefore, no load-bearing step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The platform leverages standard methods; most parameters are RL/tracking hyperparameters. No free parameter is fitted to make a prediction; the results are empirical demos. The main assumptions are LIPM planning, simulator fidelity, and the module-reuse design analogy.

free parameters (6)
  • Action scale s_j = 0.25 (all controlled lower-body joints)
    Chosen in Appendix III (Eq. 3) to map policy actions to joint offsets; a hand-set constant that shapes the locomotion controllers.
  • Tracking-reward kernel widths = 0.03 m, 0.15 rad, 0.03 m, 0.15 rad, 0.25 m/s, 0.60 rad/s
    Hand-set widths in Eq. 1 (Appendix III) controlling sensitivity of root/body position, orientation, linear/angular velocity tracking terms.
  • Velocity-policy weights and kernel widths = w_v=w_omega=2, sigma_v=0.16 m/s, sigma_omega=0.50 rad/s
    Hand-set values in Eq. 2 (Appendix III) for the reference-free velocity control objective.
  • ZMP preview LQR weights Q_y, R = not specified
    Appendix III states LQR preview control balances ZMP tracking (Q_y) vs acceleration effort (R); values are not reported, so the planner's behavior is under-specified.
  • Gait parameters T_cyc and rho = user-specified
    Gait cycle duration and single/double-support ratio in Appendix III; user-chosen per walk, not derived.
  • Diffusion policy chunk sizes = n_obs=2, n_action=8
    Appendix II; architecture hyperparameters that affect policy behavior, reported but not justified.
axioms (4)
  • domain assumption Fixed-height LIPM with constant CoM height h_com (Eq. 4: p_zmp = c_xy - (h_com/g) c_ddot_xy) governs planar walking dynamics for all generated gaits.
    Standard walking-pattern assumption (Kajita et al. [71]) but an idealization; actual CoM height varies during squat/step motions, which the planner ignores.
  • domain assumption MuJoCo and IsaacLab simulations are faithful enough that policies trained (and evaluated) there are evidence of real-world capability.
    RL tracking and velocity policies are never executed on the physical humanoid; real-world locomotion is open-loop keyframe playback. The paper's validation of 'reinforcement-learning-based locomotion' depends on this transfer assumption.
  • domain assumption The palm-to-finger vs torso-to-limb analogy justifies reusing identical kinematic modules as fingers or limbs without per-modality redesign.
    Section 3.1 states the analogy as Handroid's central design principle; if the shared modules underperform in one role (e.g., fingertip forces vs leg torques), the dual-embodiment claim weakens.
  • domain assumption Electromagnetic flange holding force (~180 N) is sufficient for manipulation and repeated docking.
    Appendix I; stated as sufficient, no margin/load analysis given.

pith-pipeline@v1.3.0-alltime-deepseek · 17338 in / 14494 out tokens · 123893 ms · 2026-08-01T21:04:49.352079+00:00 · methodology

0 comments
read the original abstract

Dexterous hands and humanoid robots are typically developed as distinct embodiments: the former enable contact-rich manipulation at the object scale, whereas the latter provide mobility and whole-body interaction in human-centered environments. We introduce \textbf{Handroid}, a desktop-scale dual-embodiment robot that integrates both capabilities within a single reconfigurable platform. Handroid reuses one 27-DoF electromechanical body as either a dexterous hand or a desktop humanoid, measuring 0.33 m in height and 2.05 kg in weight. In the dexterous hand embodiment, 20 DoFs form an anthropomorphic hand closely matching the kinematic structure of the human hand. In the humanoid embodiment, the same articulated modules are reconfigured into a humanoid with a head, arms, and legs, including a 12-DoF lower-limb structure for locomotion and whole-body motion. Handroid further provides a unified control and learning framework supporting hand teleoperation, dexterous grasping, in-hand manipulation, humanoid locomotion, gait generation, and interactive motion authoring. We validate the platform through real-world dexterous manipulation, reinforcement-learning-based locomotion, keyframe motion deployment, and a long-horizon task involving embodiment reconfiguration, locomotion, docking, and dexterous pick-and-place. These results position Handroid as a compact and reproducible platform for advancing morphology-reconfigurable robotics and cross-embodiment robot learning.

Figures

Figures reproduced from arXiv: 2607.16187 by Chenyang Ma, C. Karen Liu, Haochen Shi, Mingyu Ding, Ruogu Li, Shuran Song, Sikai Li, Yunchao Yao, Zhenyu Wei.

Figure 1
Figure 1. Figure 1: Left: Handroid demonstrates dexterous manipulation in the hand embodiment and locomotion and loco-manipulation in the humanoid embodiment. Right: Embodiment switching between the dexterous hand and humanoid morphologies. Abstract: Dexterous hands and humanoid robots are typically developed as distinct embodiments: the former enable contact-rich manipulation at the ob￾ject scale, whereas the latter provide … view at source ↗
Figure 2
Figure 2. Figure 2: Handroid Structure Design. Modules I–V correspond to the five fingers in the hand embodiment and to the head, left arm, left leg, right leg, and right arm in the humanoid embodiment, while modules VI and VII form the base and hip. We show the module mapping, joint numbering, actuator layout, and pitch/roll/yaw axes. Joints 9 and 26 drive the prismatic sliding mechanism for embodiment switching. 3 Design 3.… view at source ↗
Figure 3
Figure 3. Figure 3: Electrical Design. A compact mainboard in￾tegrates control, sensing, power management, and sta￾tus monitoring for Handroid’s dual-embodiment opera￾tion. Handroid’s electrical system is designed as a compact integrated backbone for the dual￾embodiment platform, unifying actuation con￾trol, power delivery, wireless communication, onboard sensing, and hardware-state monitor￾ing within the limited internal vol… view at source ↗
Figure 4
Figure 4. Figure 4: Teleoperation as Dexterous Hand of Handroid on various dexterous tasks. 4.1 Dexterous Hand Teleoperation. For the Dexterous Hand embodiment, we build an Apple Vision Pro-based tele￾operation interface for collecting coordinated arm-hand demonstrations. The interface retargets the operator’s hand motion to Handroid and maps wrist motion to the end-effector command of a Franka Research 3 arm, allowing the ex… view at source ↗
Figure 5
Figure 5. Figure 5: Representative Diffusion Policy grasping rollouts, conditioned on object point clouds and propri￾oception across randomized object poses [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-world cube reorientation by Handroid in the Dexterous-Hand embodiment. 1 1 2 3 4 5 6 7 8 9 10 2 3 4 5 1 2 3 4 5 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Tasks in the Humanoid embodiment, including pick-and-place, pull-up, and push-up. 5.1 Dexterous Hand Teleoperation. We qualitatively evaluate the VR-based teleoperation interface through a set of real-world manipulation tasks, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Long-horizon task with Handroid, including embodiment switching, detachment from the Franka, obstacle-avoidance walking, box pushing, lying down, re-docking with the Franka end effector, object grasping, and pick-and-place in the Dexterous-Hand embodiment. In-hand Reorientation. In addition to imitation learning, we also evaluate the in-hand reorientation policy on the real Handroid system. The Handroid in… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 6 canonical work pages

  1. [1]

    G. B. West, J. H. Brown, and B. J. Enquist. A general model for the origin of allometric scaling laws in biology.Science, 276(5309):122–126, 1997. doi:10.1126/science.276.5309.122

    work page doi:10.1126/science.276.5309.122 1997

  2. [2]

    A. A. Biewener. Biomechanical consequences of scaling.Journal of Experimental Biology, 208(9):1665–1676, 2005. doi:10.1242/jeb.01520

    work page doi:10.1242/jeb.01520 2005

  3. [3]

    Z. Wei, Z. Xu, J. Guo, Y . Hou, C. Gao, Z. Cai, J. Luo, and L. Shao.D(R,O)grasp: A unified representation of robot and object interaction for cross-embodiment dexterous grasping. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4982–4988,

    2025

  4. [4]

    J. R. Napier. The prehensile movements of the human hand.The Journal of Bone and Joint Surgery. British Volume, 38-B(4):902–913, 1956. doi:10.1302/0301-620X.38B4.902

    work page doi:10.1302/0301-620x.38b4.902 1956

  5. [5]

    J. M. F. Landsmeer. Power grip and precision handling.Annals of the Rheumatic Diseases, 21 (2):164–170, 1962. doi:10.1136/ard.21.2.164

    work page doi:10.1136/ard.21.2.164 1962

  6. [6]

    M. W. Marzke. Precision grips, hand morphology, and tools.American Journal of Phys- ical Anthropology, 102(1):91–110, 1997. doi:10.1002/(SICI)1096-8644(199701)102:1⟨91:: AID-AJPA8⟩3.0.CO;2-G

    work page doi:10.1002/(sici)1096-8644(199701)102:1 1997

  7. [7]

    Santello, M

    M. Santello, M. Flanders, and J. F. Soechting. Postural hand synergies for tool use.Journal of Neuroscience, 18(23):10105–10115, 1998. doi:10.1523/JNEUROSCI.18-23-10105.1998

    arXiv 1998

  8. [8]

    Z. Wei, Y . Yao, and M. Ding. One hand to rule them all: Canonical representations for unified dexterous manipulation.arXiv preprint arXiv:2602.16712, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills.ACM Transactions on Graphics, 37 (4):1–14, 2018

    2018

  10. [10]

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. InACM Transactions on Graphics, volume 40, 2021

    2021

  11. [11]

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi. Learning human-to-humanoid real-time whole-body teleoperation. InIEEE/RSJ International Conference on Intelligent Robots and Systems, 2024

    2024

  12. [12]

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi. Om- nih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. InConference on Robot Learning, 2024

    2024

  13. [13]

    L. Yang, X. Huang, Z. Wu, A. Kanazawa, P. Abbeel, C. Sferrazza, C. K. Liu, R. Duan, and G. Shi. Omniretarget: Interaction-preserving data generation for humanoid whole-body loco- manipulation and scene interaction, 2025. URLhttps://arxiv.org/abs/2509.26633

    Pith/arXiv arXiv 2025

  14. [14]

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, et al. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills, 2025. URL https://arxiv.org/abs/2502.01143

    Pith/arXiv arXiv 2025

  15. [15]

    Shadow dexterous hand series.https://shadowrobot.com/ dexterous-hand-series/

    Shadow Robot Company. Shadow dexterous hand series.https://shadowrobot.com/ dexterous-hand-series/. Accessed: 2026-05-21

    2026

  16. [16]

    Kumar, Y

    V . Kumar, Y . Tassa, T. Erez, and E. Todorov. Real-time behaviour synthesis for dynamic hand- manipulation.2014 IEEE International Conference on Robotics and Automation (ICRA), pages 6808–6815, 2014. URLhttps://api.semanticscholar.org/CorpusID:12906683. 16

    2014

  17. [17]

    Xynova flex 1.https://www.xynova.com.cn/en/xynova-flex-1, 2026

    Xynova. Xynova flex 1.https://www.xynova.com.cn/en/xynova-flex-1, 2026. Ac- cessed: 2026-06-25

    2026

  18. [18]

    Wuji hand 2 product introduction.https://docs.wuji.tech/docs/en/ wuji-hand/latest/overview/, 2026

    Wuji Technology. Wuji hand 2 product introduction.https://docs.wuji.tech/docs/en/ wuji-hand/latest/overview/, 2026. Accessed: 2026-06-25

    2026

  19. [19]

    Sharpa wave.https://www.sharpa.com/pages/wave

    Sharpa. Sharpa wave.https://www.sharpa.com/pages/wave. Accessed: 2026-05-21

    2026

  20. [20]

    K. Shaw, A. Agarwal, and D. Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning.Robotics: Science and Systems (RSS), 2023

    2023

  21. [21]

    Bhirangi, A

    R. Bhirangi, A. DeFranco, J. Adkins, C. Majidi, A. Gupta, T. Hellebrekers, and V . Kumar. All the feels: A dexterous hand with large-area tactile sensing, 2024. URLhttps://arxiv.org/ abs/2210.15658

    Pith/arXiv arXiv 2024

  22. [22]

    B. A. Richardson, F. Gr ¨uninger, L. Mack, J. Stueckler, and K. J. Kuchenbecker. Isyhand: A dexterous multi-finger robot hand with an articulated palm, 2025. URLhttps://arxiv. org/abs/2509.26236

    arXiv 2025

  23. [23]

    C. C. Christoph, M. Eberlein, F. Katsimalis, A. Roberti, A. Sympetheros, M. R. V ogt, D. Li- conti, C. Yang, B. G. Cangan, R. J. Hinchet, and R. K. Katzschmann. Orca: An open-source, reliable, cost-effective, anthropomorphic robotic hand for uninterrupted dexterous task learn- ing. InProceedings of the IEEE/RSJ International Conference on Intelligent Robo...

    2025

  24. [24]

    Unitree H1 - Unitree’s First Universal Humanoid Robot - Unitree, 2024

    Unitree. Unitree H1 - Unitree’s First Universal Humanoid Robot - Unitree, 2024

    2024

  25. [25]

    Unitree h2: Destiny awakening.https://www.unitree.com/H2, 2026

    Unitree Robotics. Unitree h2: Destiny awakening.https://www.unitree.com/H2, 2026. Accessed: 2026-06-25

    2026

  26. [26]

    Unitree G1 - Humanoid Agent AI Avatar - Unitree, 2024

    Unitree. Unitree G1 - Humanoid Agent AI Avatar - Unitree, 2024

    2024

  27. [27]

    Figure, 2024

    FigureAI. Figure, 2024

    2024

  28. [28]

    Dynamics

    B. Dynamics. Atlas, 2024

    2024

  29. [29]

    AI & Robotics, 2024

    Tesla. AI & Robotics, 2024

    2024

  30. [30]

    Robotics

    A. Robotics. Agility Robotics, 2024

    2024

  31. [31]

    FOURIER-Robotics, 2024

    FOURIER-Robotics. FOURIER-Robotics, 2024

    2024

  32. [32]

    Robotics

    B. Robotics. Booster, 2024

    2024

  33. [33]

    Q. Liao, B. Zhang, X. Huang, X. Huang, Z. Li, and K. Sreenath. Berkeley humanoid: A research platform for learning-based control, 2024. URLhttps://arxiv.org/abs/2407. 21781

    2024

  34. [34]

    Y . Chi, Q. Liao, J. Long, X. Huang, S. Shao, B. Nikolic, Z. Li, and K. Sreenath. Demonstrating berkeley humanoid lite: An open-source, accessible, and customizable 3d-printed humanoid robot, 2025. URLhttps://arxiv.org/abs/2504.17249

    Pith/arXiv arXiv 2025

  35. [35]

    Parmiggiani, M

    A. Parmiggiani, M. Maggiali, L. Natale, F. Nori, A. Schmitz, N. Tsagarakis, J. Santos-Victor, F. Becchi, G. Sandini, and G. Metta. The design of the icub humanoid robot.International Journal of Humanoid Robotics, 9, 12 2012. doi:10.1142/S0219843612500272

    work page doi:10.1142/s0219843612500272 2012

  36. [36]

    B. Xia, B. Li, J. Lee, M. Scutari, and B. Chen. The duke humanoid: Design and control for energy efficient bipedal locomotion using passive dynamics, 2025. URLhttps://arxiv. org/abs/2409.19795. 17

    Pith/arXiv arXiv 2025

  37. [37]

    Y . Liu, J. Shen, J. Zhang, X. Zhang, T. Zhu, and D. Hong. Design and control of a miniature bipedal robot with proprioceptive actuation for dynamic behaviors. In2022 In- ternational Conference on Robotics and Automation (ICRA), pages 8547–8553, 2022. doi: 10.1109/ICRA46639.2022.9811790

    arXiv 2022

  38. [38]

    NAO H25 — Aldebaran 2.1.4.13 Documentation, 2018

    NAO H25. NAO H25 — Aldebaran 2.1.4.13 Documentation, 2018

    2018

  39. [39]

    OP3 E-Manual, 2018

    ROBOTIS. OP3 E-Manual, 2018

    2018

  40. [40]

    K. Lab. Kscales Lab Zeroth Bot, 2025

    2025

  41. [41]

    H. Shi, W. Wang, S. Song, and K. Liu. ToddlerBot: Open-Source ML-Compatible Humanoid Platform for Loco-Manipulation. In9th Annual Conference on Robot Learning, Sept. 2025

    2025

  42. [42]

    Y . Yao, Z. Xu, T. Zhang, Z. Liu, S. Li, Z. Wei, F. Chen, D. Huang, K. Wan, C. Ma, et al. Dexverse: A modular benchmark for multi-task, multi-embodiment dexterous manipulation. arXiv preprint arXiv:2607.08751, 2026

    Pith/arXiv arXiv 2026

  43. [43]

    Handa, K

    A. Handa, K. Van Wyk, W. Yang, J. Liang, Y .-W. Chao, Q. Wan, S. Birchfield, N. Ratliff, and D. Fox. Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9164–9170. IEEE, 2020

    2020

  44. [44]

    Y . Qin, W. Yang, B. Huang, K. Van Wyk, H. Su, X. Wang, Y .-W. Chao, and D. Fox. Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system. InRobotics: Science and Systems, 2023

    2023

  45. [45]

    M. Xu, H. Zhang, Y . Hou, Z. Xu, L. Fan, M. Veloso, and S. Song. Dexumi: Using human hand as the universal manipulation interface for dexterous manipulation, 2025. URLhttps: //arxiv.org/abs/2505.21864

    arXiv 2025

  46. [46]

    C. Wang, H. Shi, W. Wang, R. Zhang, L. Fei-Fei, and C. K. Liu. Dexcap: Scalable and portable mocap data collection system for dexterous manipulation.arXiv preprint arXiv:2403.07788, 2024

    Pith/arXiv arXiv 2024

  47. [47]

    S. Zhao, X. Zhu, Y . Chen, C. Li, Y . Xie, X. Zhang, M. Ding, and M. Tomizuka. Dexh2r: Task-oriented dexterous manipulation from human to robots.IEEE/ASME Transactions on Mechatronics, 2025

    2025

  48. [48]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    2025

  49. [49]

    R. Wang, J. Zhang, J. Chen, Y . Xu, P. Li, T. Liu, and H. Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In2023 IEEE Inter- national Conference on Robotics and Automation (ICRA), 2023

    2023

  50. [50]

    P. Li, T. Liu, Y . Li, Y . Geng, Y . Zhu, Y . Yang, and S. Huang. Gendexgrasp: Generalizable dex- terous grasping. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8068–8074, 2023. doi:10.1109/ICRA48891.2023.10160667

    arXiv 2023

  51. [51]

    Y . Xu, W. Wan, J. Zhang, H. Liu, Z. Shan, H. Shen, R. Wang, H. Geng, Y . Weng, J. Chen, T. Liu, L. Yi, and H. Wang. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  52. [52]

    de Bakker, J

    V . de Bakker, J. Hejna, T. G. W. Lum, O. Celik, A. Taranovic, D. Blessing, G. Neumann, J. Bohg, and D. Sadigh. Scaffolding dexterous manipulation with vision-language models,

  53. [53]

    Zhong, J

    T. Zhong, J. Buchanan, and C. Allen-Blanchette. Grasp2grasp: Vision-based dexterous grasp translation via schr¨odinger bridges, 2025. URLhttps://arxiv.org/abs/2506.02489

    arXiv 2025

  54. [54]

    URLhttps://arxiv.org/abs/2506.19212. 18

    arXiv

  55. [55]

    Liang, Y

    Z. Liang, Y . Mu, Y . Wang, T. Chen, W. Shao, W. Zhan, M. Tomizuka, P. Luo, and M. Ding. Dexhanddiff: Interaction-aware diffusion planning for adaptive dexterous manipulation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1745–1755, 2025

    2025

  56. [56]

    Huang, Z

    D. Huang, Z. Wei, Z. Xu, Y . Yao, S. Li, and M. Ding. Dexcompose: Reusing dexterous policies for multi-task manipulation with a single hand.arXiv preprint arXiv:2606.28323, 2026

    Pith/arXiv arXiv 2026

  57. [57]

    Akkaya, M

    OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng, Q. Yuan, W. Zaremba, and L. Zhang. Solving rubik’s cube with a robot hand.arXiv preprint arXiv:1910.07113, 2019

    Pith/arXiv arXiv 1910

  58. [58]

    Rajeswaran, V

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstra- tions. InRobotics: Science and Systems, 2018

    2018

  59. [59]

    H. Qi, A. Kumar, R. Calandra, Y . Ma, and J. Malik. In-hand object rotation via rapid motor adaptation. InProceedings of The 6th Conference on Robot Learning, volume 205 ofProceed- ings of Machine Learning Research, pages 1722–1732, 2023

    2023

  60. [60]

    Andrychowicz, B

    OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba. Learning dexterous in-hand manipulation.The International Jour- nal of Robotics Research, 39(1):3–20, 2020. doi:10.1177/0278364919887447

    doi:10.1177/0278364919887447 2020

  61. [61]

    C. Ma, Y . Yao, Z. Wei, R. Li, D. Szafir, and M. Ding. Current as touch: Proprioceptive contact feedback for compliant dexterous manipulation.arXiv preprint arXiv:2607.03529, 2026

    Pith/arXiv arXiv 2026

  62. [62]

    H. Qi, B. Yi, S. Suresh, M. Lambeta, Y . Ma, R. Calandra, and J. Malik. General in-hand object rotation with vision and touch. InConference on Robot Learning, 2023

    2023

  63. [63]

    T. Chen, M. Tippur, S. Wu, V . Kumar, E. Adelson, and P. Agrawal. Visual dexterity: In-hand reorientation of novel and complex object shapes.Science Robotics, 8(84):eadc9244, 2023. doi:10.1126/scirobotics.adc9244

    doi:10.1126/scirobotics.adc9244 2023

  64. [64]

    Handa, A

    A. Handa, A. Allshire, V . Makoviychuk, A. Petrenko, R. Singh, J. Liu, D. Makoviichuk, K. Van Wyk, A. Zhurkevich, B. Sundaralingam, Y . Narang, J.-F. Lafleche, D. Fox, and G. State. Dextreme: Transfer of agile in-hand manipulation from simulation to reality. In2023 IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  65. [65]

    Gu, Y .-J

    X. Gu, Y .-J. Wang, and J. Chen. Humanoid-gym: Reinforcement learning for humanoid robot with zero-shot sim2real transfer.arXiv preprint arXiv:2404.05695, 2024

    Pith/arXiv arXiv 2024

  66. [66]

    T. Lin, K. Sachdev, L. Fan, J. Malik, and Y . Zhu. Sim-to-real reinforcement learning for vision- based dexterous manipulation on humanoids, 2025. URLhttps://arxiv.org/abs/2502. 20396

    2025

  67. [67]

    K. Hu, H. Shi, Y . He, W. Wang, C. K. Liu, and S. Song. Robot trains robot: Automatic real-world policy adaptation and learning for humanoids. InConference on Robot Learning, 2025. 19

    2025

  68. [68]

    Zhang, Y

    Y . Zhang, Y . Yuan, P. Gurunath, I. Gupta, S. Omidshafiei, A.-a. Agha-mohammadi, M. Vazquez-Chanlatte, L. Pedersen, T. He, and G. Shi. Falcon: Learning force-adaptive hu- manoid loco-manipulation, 2025. URLhttps://arxiv.org/abs/2505.06776

    arXiv 2025

  69. [69]

    S. Li, S. Li, J. Li, and M. Ding. Anybody: Free-form whole-body humanoid control from arbitrary keypoint guidance, 2026. URLhttps://arxiv.org/abs/2606.29209

    Pith/arXiv arXiv 2026

  70. [70]

    S. Li, S. Li, Z. Wei, Y . Yao, C. Li, and M. Ding. Coordex: Coordinating body and hand pri- ors for continuous dexterous humanoid loco-manipulation.arXiv preprint arXiv:2606.23680, 2026

    Pith/arXiv arXiv 2026

  71. [71]

    Kajita, F

    S. Kajita, F. Kanehiro, K. Kaneko, K. Fujiwara, K. Harada, K. Yokoi, and H. Hirukawa. Biped walking pattern generation by using preview control of zero-moment point. InProceedings of the 2003 IEEE International Conference on Robotics and Automation, pages 1620–1626. IEEE, 2003

    2003

  72. [72]

    R. Li, S. Li, Y . Mu, and M. Ding. Sldprtnet: A large-scale multimodal dataset for cad genera- tion in language-driven 3d design, 2026. URLhttps://arxiv.org/abs/2603.13098

    arXiv 2026

  73. [73]

    T. H. Yang, H. Shi, J. Hu, Z. Zhang, D. Jiang, W. Wang, Y . He, Z. Wu, Y . Chen, Y . Hou, et al. Locomotion beyond feet.arXiv preprint arXiv:2601.03607, 2026

    arXiv 2026

  74. [74]

    K. Zakka. Mink: Python inverse kinematics based on MuJoCo, Feb. 2026. URLhttps: //github.com/kevinzakka/mink

    2026

  75. [75]

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    2017

  76. [76]

    B. Wen, W. Yang, J. Kautz, and S. Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17868–17879, 2024

    2024

  77. [77]

    Mittal, P

    M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Mu ˜noz, X. Yao, R. Zurbr ¨ugg, N. Rudin, L. Wawrzyniak, M. Rakhsha, A. Denzler, E. Heiden, A. Borovicka, O. Ahmed, I. Akinola, A. Anwar, M. T. Carlson, J. Y . Feng, A. Garg, R. Gasoto, L. Gulich, Y . Guo, M. Gussert, A. Hansen, M. Kulkarni, C. Li, W. Liu, V . Makoviychuk, G. Malczyk, H...

    Pith/arXiv arXiv 2025

  78. [78]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical im- age segmentation. InInternational Conference on Medical image computing and computer- assisted intervention, pages 234–241. Springer, 2015

    2015

  79. [80]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017. 20

    Pith/arXiv arXiv 2017

  80. [2025]

    doi:10.1109/ICRA55743.2025.11127754

    arXiv 2025