Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

The paper reports that a 35 kg humanoid, trained to lean its torso from a static torque-balance reward, pulls 168 N (48% of its body weight), about 69% above the best baseline, and transfers to real fire-door and rack-pulling tasks.

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 ↗

A decoupled whole-body RL policy with a force-based lean reward enables a Unitree G1 humanoid to pull with up to 167.7 N, beating prior controllers by 69–75%.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Real empirical gains on a G1 humanoid from a torso-tilt reward and waist-decoupled policy, but the unverified Falcon baseline and a hand-wavy torque-balance derivation keep the headline numbers from being decisive. the 4 major comments →

arxiv 2510.26280 v3 pith:2PSJH2EA submitted 2025-10-30 cs.RO

Thor: Towards Human-Level Whole-Body Reactions for Intense Contact-Rich Environments

Gangyang Li , Qing Shi , Youhao Hu , Zhongyuan Wang , Xinlong Wang , Shaqi Luo This is my paper
classification cs.RO
keywords humanoid whole-body controlforce-adaptive torso-tilt rewardreinforcement learningcontact-rich manipulationzero-moment-point balanceloco-manipulationsim-to-real policy transferforce interaction
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.

The reading

The paper's claim is that human-like torso leaning is the main thing separating a humanoid that strains against a load from one that actually moves it. The authors derive a force-adaptive torso-tilt (FAT2) reward from a zero-moment-point torque-balance analysis: for a given hand force, there is a target lean angle that keeps the ground-reaction pivot inside the foot support polygon while letting gravity contribute to the pull. They train three decoupled policies — upper body, waist, lower body — with shared whole-body observations, then deploy the result on a 1.32 m, 35 kg humanoid. In repeated real-world measurements the robot's peak backward pull is 167.7 ± 2.4 N (about 48% of body weight) and forward pull is 145.5 ± 2.0 N, corresponding to 68.9% and 74.7% improvements over the best baseline, and it opens a fire door with one hand and pulls a rack carrying 70 kg. The paper also notes that the multi-agent setup still depends on manual tuning of reward scales and learning rates.

Core claim

On the paper's own terms, the central discovery is that a quasi-static torque-balance model of a humanoid under an external end-effector force produces a simple closed-form target lean angle, and using that target as a reward term makes the learned policy lean with the force — in some states carrying the center of mass outside the foot support polygon — and that this raises peak pull force from about 99 N to about 168 N relative to the best baseline. The ablation attributes roughly 80–90% of the full performance to the FAT2 reward alone, while the decoupled upper/waist/lower architecture removes residual artifacts, such as the waist roll deviations seen under heavy pulls. At deployment no fo

What carries the argument

The load-bearing mechanism is the FAT2 reward, exp(−||β−β′||²/σ_t), where β is the current torso pitch and β′ is a target computed from the scalar torque balance β′ = arccos( |F_h||r_h|cosφ cosα / (|F_g||r_CoM|) ), capped at 0.9 rad. The formula encodes the zero-moment-point (ZMP) balance criterion — the ground-reaction pivot must stay inside the foot support polygon — as a reward; the hand's horizontal offset d3 is dropped because its angle is considered small. The second mechanism is the decoupled policy architecture: three independent actor-critic networks for the upper body (14 joints), waist (3 joints), and lower body (12 joints), each receiving shared whole-body observations, trained j

Load-bearing premise

The load-bearing premise is that the robot's motion is slow enough for a static torque balance to give the right lean angle — the derivation drops the hand's horizontal offset and ignores acceleration — so if real pulls are fast or offset, the predicted lean angle could be miscalibrated and the reported force gains would be reward-shaping effects rather than the claimed biomechanical mechanism.

What would settle it

Measure the robot's actual torso pitch at the moment of peak pulling force across handle heights and pull speeds, and compare it with the torque-balance target predicted by Eq. (17); a large systematic mismatch under dynamic or offset pulls would indicate that the learned leaning is not tracking the proposed mechanism. A complementary test is to retrain with the dropped d3 term and nonzero acceleration terms included: if the force gains disappear or the lean angle shifts, the quasi-static simplification was load-bearing.

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

If this is right

  • A humanoid can pull about 48% of its own body weight while moving backward and 42% while moving forward, roughly 69% and 75% above the best baseline in the paper.
  • The torque-balance lean target can be used as a training reward without force sensors at deployment; the network learns to read the force from the torso's pitch angle and pitch velocity.
  • Controlled placement of the center of mass outside the support polygon is a viable learned strategy for generating high interaction force, rather than a failure mode.
  • The decoupled upper/waist/lower architecture runs at 50 Hz on limited onboard compute while retaining whole-body coordination through shared observations.
  • The policy transfers to practical tasks: one-hand fire-door opening at about 60 N and pulling a rack loaded with 70 kg at about 130 N, where the baselines fail.

Where Pith is reading between the lines

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

  • Beyond the paper: if the closed-form lean target is as predictive as reported, the same torque-balance equation could be inverted into a feedforward lean-angle reference for model-based controllers, potentially giving comparable force gains without reinforcement learning.
  • Beyond the paper: because the waist network emerges as the force-transmission layer, a natural test is to fix the waist and lower-body policies and retrain only the upper-body policy for new manipulation tasks, which would make the framework modular.
  • Beyond the paper: the quasi-static derivation drops the hand's horizontal offset and assumes negligible acceleration; a direct extension would add those terms and test whether the learned lean angle tracks the corrected target under fast pulls and off-center handle heights.
  • Beyond the paper: the force gains generalize only if ground friction is sufficient; the paper uses custom shoe covers to ensure friction, so an open question is how the lean strategy degrades on low-friction floors.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes Thor, a reinforcement-learning framework for humanoid whole-body control in forceful interaction tasks. The method has two main ingredients: (1) a decoupled policy architecture in which separate actor-critic networks control the upper body, waist, and lower body while sharing whole-body observations, and (2) a force-adaptive torso-tilt (FAT2) reward derived from a quasi-static ZMP/torque-balance analysis, which encourages the robot to lean its torso in response to applied end-effector forces. The policy is trained in Isaac Gym with curriculum learning and domain randomization, then deployed on a Unitree G1. Real-world measurements report peak pulling forces of 167.7±2.4 N backward and 145.5±2.0 N forward, improvements of 68.9% and 74.7% over the Falcon baseline, and successful fire-door and loaded-rack tasks. Ablations indicate that FAT2 contributes most of the performance gain, with the decoupled architecture providing additional stabilization.

Significance. If the empirical results hold, the paper demonstrates a practical recipe for substantially increasing humanoid interaction force using reward shaping derived from a simple static model, combined with a modular policy architecture. The work includes real-hardware validation, quantitative force measurements, ablations, and a public project page—strengths that make the core claims testable. However, the force improvements are relative to a Falcon baseline whose training setup is not described, and the FAT2 derivation contains a non-trivial simplification that is not quantitatively justified. These issues do not invalidate the measured improvements, but they affect the strength of the comparative claim and the explanatory narrative. The contribution is potentially useful to the humanoid-locomotion and loco-manipulation community, provided the baseline and derivation concerns are resolved.

major comments (4)
  1. [§III-B, Eqs. (15)–(17)] The derivation drops the term d3=|rh|sinφ with the statement 'sinφ is very small,' but no numerical bound is given for the actual postures used in training or deployment. In the reported pulling tasks, the hands are not directly above the support polygon; d3 can be comparable to d1·cosα. Since Eq. (17) is used to compute the FAT2 lean target, an unjustified neglect of d3 can miscalibrate the reward. Please provide a quantitative comparison of d3 and d1·cosα for the G1 postures in Fig. 4/5, or otherwise justify the simplification. Also justify the quasi-static assumption (negligible acceleration) for the dynamic pulling motions shown in Fig. 5.
  2. [§III-B, Eq. (18) and Table I] The FAT2 reward is written as exp(−‖β−β′‖²/σt), but β′ is never defined, and the width σt is not reported in Table I or elsewhere. More broadly, the complete reward function—including terms for tracking, stability, and the individual rewards for the upper/waist/lower agents—is not specified. Without these details, the method cannot be reproduced, and it is difficult to verify that the force gains are attributable to FAT2 rather than to other reward terms or reward scales. Please report the full reward specification and all associated constants.
  3. [§IV-B, Table II] Falcon is the strongest baseline and the reference for the headline 68.9%/74.7% improvements, but the manuscript gives no training details for this baseline: whether the authors' official implementation/checkpoints were used, the hyperparameters, number of seeds, or domain-randomization settings. If Falcon was re-implemented or under-trained, the reported gap could be inflated. Please include a complete description of the baseline training protocol, or use the official Falcon checkpoints and report their performance. This is essential for the claim of 'consistently outperforms the baseline algorithms.'
  4. [§IV-B, Table II] Each entry is reported as mean ± SE, but the number of repeated trials per condition is not stated, and no statistical test is performed. Without trial counts and significance testing, the claim of consistent outperformance is not fully supported—particularly for cells where Thor and Falcon overlap (e.g., F_0°_sf and F_180°_dp). Please provide the number of trials per condition and, if feasible, a pairwise test or confidence intervals.
minor comments (5)
  1. [§III-Fig. 3] The symbols α, φ, d1, d2, d3 are introduced in the figure but α and φ are not explicitly defined in the caption or text; please clarify their geometric meaning and label the coordinate axes.
  2. [General] There are several typos and grammatical issues: 'simulatior' (Isaac Gym), 'the its body weight' in the abstract, 'the robot’s CoM is completely located outside' (suggest rephrasing). A careful copyedit is recommended.
  3. [§II-A] The statement that prior methods 'typically assume that the robot’s CoM projection lies within the support region' is not supported by direct citations for the mentioned baselines; please clarify which specific methods make this assumption and what evidence supports the claim that Thor violates it (e.g., CoM tracking data).
  4. [§IV-B] The custom shoe covers used to increase ground friction are mentioned for real-world scenarios, but the friction coefficient is not reported. Since friction is critical for pulling tasks, please quantify the modification.
  5. [Eq. (8)] The term C(a_i^t) is called 'torque regularization,' but it is an L2 penalty on the joint-position actions, not on torques. Please rename or clarify to avoid confusion.

Circularity Check

0 steps flagged

No significant circularity; the FAT2 reward is a design choice and the headline force results are external measurements.

full rationale

Thor's derivation chain is not circular. FAT2 (Eqs. 15-18) is constructed from a quasi-static ZMP torque balance to define a target torso tilt angle beta' as a function of the privileged interaction force, and is then used as a reward-shaping term. This is a design input, not a fitted parameter disguised as a prediction: the paper explicitly states that FAT2 encourages the tilting behavior, so the observed torso-lean is not claimed as an independent discovery. The central empirical claims (Table II peak forces of 167.7±2.4 N and 145.5±2.0 N; fire-door and loaded-rack tasks) are measured in simulation and on hardware against Falcon, Homie, and Unitree default policies, and the ablations Thor1/Thor2 compare variants trained with/without FAT2 and with/without the decoupled structure. None of these comparisons is defined in terms of the FAT2 torque-balance equation. There are no load-bearing self-citations: Falcon, Homie, AMASS, and PPO are external prior works by different author groups, and no uniqueness theorem or prior result by the present authors is invoked to force the architecture or reward. The approximations noted (dropping d3=|rh|sinφ and identifying beta as both torso tilt and CoM angle) are modeling-accuracy concerns, not circular reductions. Concerns that Falcon baseline training details are unverified are reproducibility/strength-of-baseline concerns and belong under correctness risk, not circularity. Under the rule that only specific quote-and-reduction circularity counts, the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central method rests on one hand-set cap, several unreported reward/distribution constants, and an approximate torque balance; no new physical entities are introduced.

free parameters (3)
  • β_max (maximum torso lean) = 0.9 rad
    Set 'empirically' in §III-B; caps the lean angle in Eqs. 17–19 and thereby the force the reward aims for; no sensitivity analysis is provided.
  • σ_t (FAT2 Gaussian width) = not reported
    Eq. 18 uses exp(-||β-β'||²/σ_t); the width and the associated reward weight are omitted, and the conclusion admits reward scaling factors require manual tuning.
  • Domain-randomization force distribution = not reported
    Gaussian perturbations are 'applied to the direction and magnitude of forces' in §III-A; mean and std are not specified, yet they determine the sim-to-real transfer.
axioms (4)
  • domain assumption Quasi-static rigid-body equilibrium
    §III-B: force/torque equilibrium with negligible acceleration is the basis for Eqs. 10–17; not validated for the dynamic forward/backward tasks.
  • ad hoc to paper Hand horizontal offset d3 is negligible
    Eq. 16 drops |Fh| d3 sinα because 'sinφ being very small'; no bound is given, and the term appears in Eq. 15 for realistic reaching postures.
  • domain assumption Human-like torso lean improves force transmission
    §III-B cites biomechanics refs [17],[18] to justify FAT2, but no G1-specific biomechanical validation is provided.
  • domain assumption ZMP criterion is sufficient for balance under external force
    Used to frame torque balance about the support-polygon centroid; ignores foot compliance, ankle torques, and multi-contact effects.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Thor: Towards Human-Level Whole-Body Reactions for Intense Contact-Rich Environments." pith.science (2026). https://pith.science/paper/2PSJH2EA

@misc{pith2026251026280,
  author       = {Pith},
  title        = {Pith review of: Thor: Towards Human-Level Whole-Body Reactions for Intense Contact-Rich Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PSJH2EA}},
  note         = {Machine review of arXiv:2510.26280}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Humanoids hold great potential for service, industrial, and rescue applications, in which robots must sustain whole-body stability while performing intense, contact-rich interactions with the environment. However, enabling humanoids to generate human-like, adaptive responses under such conditions remains a major challenge. To address this, we propose Thor, a humanoid framework for human-level whole-body reactions in contact-rich environments. Based on the robot's force analysis, we design a force-adaptive torso-tilt (FAT2) reward function to encourage humanoids to exhibit human-like responses during force-interaction tasks. To mitigate the high-dimensional challenges of humanoid control, Thor introduces a reinforcement learning architecture that decouples the upper body, waist, and lower body. Each component shares global observations of the whole body and jointly updates its parameters. Finally, we deploy Thor on the Unitree G1, and it substantially outperforms baselines in force-interaction tasks. Specifically, the robot achieves a peak pulling force of 167.7 N (approximately 48% of the G1's body weight) when moving backward and 145.5 N when moving forward, representing improvements of 68.9% and 74.7%, respectively, compared with the best-performing baseline. Moreover, Thor is capable of pulling a loaded rack (130 N) and opening a fire door with one hand (60 N). These results highlight Thor's effectiveness in enhancing humanoid force-interaction capabilities.

Figures

Figures reproduced from arXiv: 2510.26280 by Gangyang Li, Qing Shi, Shaqi Luo, Xinlong Wang, Youhao Hu, Zhongyuan Wang.

Figure 1
Figure 1. Figure 1: Humanoids performing tasks involving forceful interactions with the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of Thor. The whole-body control strategy for humanoids is decoupled into a network architecture comprising the upper body, waist, and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Humanoid force interaction analysis with ZMP constraint. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sequential plots of the robot’s posture and the corresponding interactive force in the simulation environment: (a) backward motion, (b) forward [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The variation of the pulling force generated by the robot with [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. BifrostUMI: Bridging Robot-Free Demonstrations and Humanoid Whole-Body Manipulation

    cs.RO 2026-05 unverdicted novelty 6.0

    BifrostUMI enables robot-free human demonstration capture via VR and wrist cameras to train visuomotor policies that predict keypoint trajectories for transfer to humanoid whole-body control through retargeting.

Reference graph

Works this paper leans on

48 extracted references · 16 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Hold my beer: Learning gentle humanoid locomotion and end-effector stabilization control,

    Y . Li, Y . Zhang, W. Xiao, C. Pan, H. Weng, G. He, T. He, and G. Shi, “Hold my beer: Learning gentle humanoid locomotion and end-effector stabilization control,” 2025. [Online]. Available: https://arxiv.org/abs/2505.24198

    Pith/arXiv arXiv 2025

  2. [2]

    Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,

    F. Liu, Z. Gu, Y . Cai, Z. Zhou, H. Jung, J. Jang, S. Zhao, S. Ha, Y . Chen, D. Xu, and Y . Zhao, “Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation,” 2025. [Online]. Available: https://arxiv.org/abs/2409.20514

    arXiv 2025

  3. [3]

    Motion generation for pulling a fire hose by a humanoid robot,

    I. G. Ramirez-Alpizar, M. Naveau, C. Benazeth, O. Stasse, J.-P. Laumond, K. Harada, and E. Yoshida, “Motion generation for pulling a fire hose by a humanoid robot,” in2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids), 2016, pp. 1016–1021

    2016

  4. [4]

    Adaptive-force-based control of dynamic legged locomotion over uneven terrain,

    M. Sombolestan and Q. Nguyen, “Adaptive-force-based control of dynamic legged locomotion over uneven terrain,”IEEE Transactions on Robotics, vol. 40, pp. 2462–2477, 2024

    2024

  5. [5]

    Humanoid loco- manipulations pattern generation and stabilization control,

    M. Murooka, K. Chappellet, A. Tanguy, M. Benallegue, I. Kuma- gai, M. Morisawa, F. Kanehiro, and A. Kheddar, “Humanoid loco- manipulations pattern generation and stabilization control,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 5597–5604, 2021

    2021

  6. [6]

    Interaction force reconstruction for humanoid robots,

    T. Mattioli and M. Vendittelli, “Interaction force reconstruction for humanoid robots,”IEEE Robotics and Automation Letters, vol. 2, no. 1, pp. 282–289, 2017

    2017

  7. [7]

    Quadratic programming for multirobot and task-space force control,

    K. Bouyarmane, K. Chappellet, J. Vaillant, and A. Kheddar, “Quadratic programming for multirobot and task-space force control,”IEEE Transactions on Robotics, vol. 35, no. 1, pp. 64–77, 2019

    2019

  8. [8]

    Torque-based balancing for a humanoid robot performing high-force interaction tasks,

    F. Abi-Farraj, B. Henze, C. Ott, P. R. Giordano, and M. A. Roa, “Torque-based balancing for a humanoid robot performing high-force interaction tasks,”IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 2023–2030, 2019

    2023

  9. [9]

    Falcon: Learning force-adaptive humanoid loco-manipulation,

    Y . Zhang, Y . Yuan, P. Gurunath, T. He, S. Omidshafiei, A. akbar Agha-mohammadi, M. Vazquez-Chanlatte, L. Pedersen, and G. Shi, “Falcon: Learning force-adaptive humanoid loco-manipulation,” 2025. [Online]. Available: https://arxiv.org/abs/2505.06776

    arXiv 2025

  10. [10]

    Sim-to-real learning for humanoid box loco-manipulation,

    J. Dao, H. Duan, and A. Fern, “Sim-to-real learning for humanoid box loco-manipulation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 16 930–16 936

    2024

  11. [11]

    Learning force con- trol for legged manipulation,

    T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal, “Learning force con- trol for legged manipulation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 15 366–15 372

    2024

  12. [12]

    Humanplus: Humanoid shadowing and imitation from humans,

    Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn, “Humanplus: Humanoid shadowing and imitation from humans,” in8th Annual Conference on Robot Learning, 2024. [Online]. Available: https://openreview.net/forum?id=WnSl42M9Z4

    2024

  13. [13]

    Mobile-television: Predictive motion priors for humanoid whole-body control,

    C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang, “Mobile-television: Predictive motion priors for humanoid whole-body control,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 5364–5371

    2025

  14. [14]

    The 3d linear inverted pendulum mode: a simple modeling for a biped walking pattern generation,

    S. Kajita, F. Kanehiro, K. Kaneko, K. Yokoi, and H. Hirukawa, “The 3d linear inverted pendulum mode: a simple modeling for a biped walking pattern generation,” inProceedings 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems. Expand- ing the Societal Role of Robotics in the the Next Millennium (Cat. No.01CH37180), vol. 1, 2001, pp. 2...

    2001

  15. [15]

    Learning advanced locomotion for quadrupedal robots: A distributed multi-agent reinforcement learning framework with riemannian motion policies,

    Y . Wang, R. Sagawa, and Y . Yoshiyasu, “Learning advanced locomotion for quadrupedal robots: A distributed multi-agent reinforcement learning framework with riemannian motion policies,”Robotics, vol. 13, no. 6, 2024. [Online]. Available: https://www.mdpi.com/2218-6581/13/6/86

    2024

  16. [16]

    Multi-agent deep reinforcement learn- ing: a survey,

    S. Gronauer and K. Diepold, “Multi-agent deep reinforcement learn- ing: a survey,”Artificial Intelligence Review, vol. 55, no. 2, pp. 895– 943, 2022

    2022

  17. [17]

    Analysis of tug of war competition: A narrative complete review,

    R. Cayero, V . Rocandio, A. Zubillaga, I. Refoyo, J. Calleja-Gonz ´alez, A. Casta ˜neda-Babarro, and I. Mart ´ınez de Aldama, “Analysis of tug of war competition: A narrative complete review,”International Journal of Environmental Research and Public Health, vol. 19, no. 1,

  18. [18]

    Low back biomechanics during repetitive deadlifts: A narrative review,

    V . J. Ramirez, B. Bazrgari, F. Gao, and M. Samaan, “Low back biomechanics during repetitive deadlifts: A narrative review,”IISE transactions on occupational ergonomics and human factors, vol. 10, no. 1, pp. 34–46, 2022

    2022

  19. [19]

    Hierarchical adaptive loco- manipulation control for quadruped robots,

    M. Sombolestan and Q. Nguyen, “Hierarchical adaptive loco- manipulation control for quadruped robots,” in2023 IEEE Interna- tional Conference on Robotics and Automation (ICRA), 2023, pp. 12 156–12 162

    2023

  20. [20]

    Hierarchical optimization-based control for whole-body loco-manipulation of heavy objects,

    A. Rigo, M. Hu, S. K. Gupta, and Q. Nguyen, “Hierarchical optimization-based control for whole-body loco-manipulation of heavy objects,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 15 322–15 328

    2024

  21. [21]

    Full-order sampling-based mpc for torque-level locomotion control via diffusion-style annealing,

    H. Xue, C. Pan, Z. Yi, G. Qu, and G. Shi, “Full-order sampling-based mpc for torque-level locomotion control via diffusion-style annealing,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 4974–4981

    2025

  22. [22]

    Dynamic loco- manipulation on hector: Humanoid for enhanced control and open- source research,

    J. Li, J. Ma, O. Kolt, M. Shah, and Q. Nguyen, “Dynamic loco- manipulation on hector: Humanoid for enhanced control and open- source research,”arXiv preprint arXiv:2312.11868, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Kinodynamics-based pose optimization for humanoid loco-manipulation,

    J. Li and Q. Nguyen, “Kinodynamics-based pose optimization for humanoid loco-manipulation,” 2023. [Online]. Available: https://arxiv.org/abs/2303.04985

    Pith/arXiv arXiv 2023

  24. [24]

    A humanoid robot carrying a heavy object,

    K. Harada, S. Kajita, H. Saito, M. Morisawa, F. Kanehiro, K. Fujiwara, K. Kaneko, and H. Hirukawa, “A humanoid robot carrying a heavy object,” inProceedings of the 2005 IEEE International Conference on Robotics and Automation, 2005, pp. 1712–1717

    2005

  25. [25]

    On the emergence of whole- body strategies from humanoid robot push-recovery learning,

    D. Ferigo, R. Camoriano, P. M. Viceconte, D. Calandriello, S. Traver- saro, L. Rosasco, and D. Pucci, “On the emergence of whole- body strategies from humanoid robot push-recovery learning,”IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 8561–8568, 2021

    2021

  26. [26]

    Wococo: Learning whole-body humanoid control with sequential contacts,

    C. Zhang, W. Xiao, T. He, and G. Shi, “Wococo: Learning whole-body humanoid control with sequential contacts,” in8th Annual Conference on Robot Learning, 2024. [Online]. Available: https://openreview.net/forum?id=Czs2xH9114

    2024

  27. [27]

    Rambo: Rl- augmented model-based whole-body control for loco-manipulation,

    J. Cheng, D. Kang, G. Fadini, G. Shi, and S. Coros, “Rambo: Rl- augmented model-based whole-body control for loco-manipulation,” IEEE Robotics and Automation Letters, vol. 10, no. 9, pp. 9462–9469, 2025

    2025

  28. [28]

    Bridging the sim-to-real gap for athletic loco-manipulation,

    N. Fey, G. B. Margolis, M. Peticco, and P. Agrawal, “Bridging the sim-to-real gap for athletic loco-manipulation,” 2025. [Online]. Available: https://arxiv.org/abs/2502.10894

    Pith/arXiv arXiv 2025

  29. [29]

    Facet: Force-adaptive control via impedance reference tracking for legged robots,

    B. Xu, H. Weng, Q. Lu, Y . Gao, and H. Xu, “Facet: Force-adaptive control via impedance reference tracking for legged robots,” 2025. [Online]. Available: https://arxiv.org/abs/2505.06883

    Pith/arXiv arXiv 2025

  30. [30]

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

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi, “Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning,”CoRR, vol. abs/2406.08858,

    Pith/arXiv arXiv

  31. [31]

    Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbabu, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. J. Fan, Y . Zhu, C. Liu, and G. Shi, “Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills,”In Robotics: Science and Systems (RSS), 2025

    2025

  32. [32]

    Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,

    Q. Liao, T. E. Truong, X. Huang, G. Tevet, K. Sreenath, and C. K. Liu, “Beyondmimic: From motion tracking to versatile humanoid control via guided diffusion,” 2025. [Online]. Available: https://arxiv.org/abs/2508.08241

    Pith/arXiv arXiv 2025

  33. [33]

    Exbody2: Advanced expressive humanoid whole-body control,

    M. Ji, X. Peng, F. Liu, J. Li, G. Yang, X. Cheng, and X. Wang, “Exbody2: Advanced expressive humanoid whole-body control,”

  34. [34]

    Visual imitation enables contextual humanoid control,

    A. Allshire, H. Choi, J. Zhang, D. McAllister, A. Zhang, C. M. Kim, T. Darrell, P. Abbeel, J. Malik, and A. Kanazawa, “Visual imitation enables contextual humanoid control,” 2025. [Online]. Available: https://arxiv.org/abs/2505.03729

    Pith/arXiv arXiv 2025

  35. [35]

    Learning human-to-humanoid real-time whole-body teleoperation,

    T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 8944–8951

    2024

  36. [36]

    Gmt: General motion tracking for humanoid whole-body control,

    Z. Chen, M. Ji, X. Cheng, X. Peng, X. B. Peng, and X. Wang, “Gmt: General motion tracking for humanoid whole-body control,” arXiv:2506.14770, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    Expressive whole-body control for humanoid robots,

    X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Expressive whole-body control for humanoid robots,”In Robotics: Science and Systems (RSS), 2024

    2024

  38. [38]

    Hover: Versatile neural whole-body controller for humanoid robots,

    T. He, W. Xiao, T. Lin, Z. Luo, Z. Xu, Z. Jiang, J. Kautz, C. Liu, G. Shi, X. Wang, L. J. Fan, and Y . Zhu, “Hover: Versatile neural whole-body controller for humanoid robots,” in2025 IEEE Inter- national Conference on Robotics and Automation (ICRA), 2025, pp. 9989–9996

    2025

  39. [39]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Ara ´ujo, Z. ang Cao, X. B. Peng, J. Wu, and C. K. Liu, “Twist: Teleoperated whole-body imitation system,” 2025. [Online]. Available: https://arxiv.org/abs/2505.02833

    Pith/arXiv arXiv 2025

  40. [40]

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

    Y . Li, Y . Lin, J. Cui, T. Liu, W. Liang, Y . Zhu, and S. Huang, “Clone: Closed-loop whole-body humanoid teleoperation for long-horizon tasks,” 2025. [Online]. Available: https://arxiv.org/abs/2506.08931

    Pith/arXiv arXiv 2025

  41. [41]

    Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit,

    Q. Ben, F. Jia, J. Zeng, J. Dong, D. Lin, and J. Pang, “Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit,”

  42. [42]

    Amo: Adaptive motion optimization for hyper- dexterous humanoid whole-body control,

    J. Li, X. Cheng, T. Huang, S. Yang, R.-Z. Qiu, and X. Wang, “Amo: Adaptive motion optimization for hyper- dexterous humanoid whole-body control,” 2025. [Online]. Available: https://arxiv.org/abs/2505.03738

    Pith/arXiv arXiv 2025

  43. [43]

    Amass: Archive of motion capture as surface shapes,

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black, “Amass: Archive of motion capture as surface shapes,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019

    2019

  44. [44]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017. [Online]. Available: https://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  45. [45]

    Available: https://arxiv.org/abs/2502.13013

    [Online]. Available: https://arxiv.org/abs/2502.13013

    Pith/arXiv arXiv

  46. [2022]

    Available: https://www.mdpi.com/1660-4601/19/1/3

    [Online]. Available: https://www.mdpi.com/1660-4601/19/1/3

  47. [2024]

    Available: https://doi.org/10.48550/arXiv.2406.08858

    [Online]. Available: https://doi.org/10.48550/arXiv.2406.08858

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.08858

  48. [2025]

    Available: https://arxiv.org/abs/2412.13196

    [Online]. Available: https://arxiv.org/abs/2412.13196

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.