Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

SkillBlender: Towards Versatile Humanoid Whole-Body Loco-Manipulation via Skill Blending

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read SkillBlender claims that blending four pretrained primitive skills under a high-level controller yields accurate, feasible humanoid loco-manipulation with one or two reward terms per task.

desk verdict Solid HRL paper with a genuinely useful benchmark; the central result mostly holds, but 'versatile' and 'significantly outperforms' outrun the evidence. read the letter →

arxiv 2506.09366 v1 pith:5H55TE7T submitted 2025-06-11 cs.RO cs.LG

Yuxuan Kuang , Haoran Geng , Amine Elhafsi , Tan-Dzung Do , Pieter Abbeel , Jitendra Malik , Marco Pavone , Yue Wang This is my paper
classification cs.ROcs.LG
keywords hierarchicalreinforcementlearninghumanoidloco-manipulationskillblendingwhole-bodycontrolpretrain-then-blendsimulationbenchmarkmotionfeasibility
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that humanoid robots can learn diverse whole-body loco-manipulation tasks without task-specific reward shaping, by first pretraining a library of task-agnostic skills and then letting a high-level controller blend them. The core claim is that this pretrain-then-blend structure, with the controller emitting subgoals and per-joint weights, outperforms both from-scratch reinforcement learning and earlier hierarchical designs on the eight tasks of the new SkillBench benchmark, and does so with one or two reward terms per task. A sympathetic reader should care because the approach targets the main bottleneck in humanoid learning: each new task currently requires tedious reward engineering, and the paper offers evidence that reusable skills can remove most of it.

What carries the argument

The central object is the vectorized skill-blending operator: a high-level policy $\pi_H$ outputs, for each selected primitive, a subgoal and a per-joint weight vector; weights are normalized by a joint-level softmax (Eq.~1), and the final action is the Hadamard-weighted sum of the primitive actions (Eq.~2). This machinery converts task solving from reward-shaped exploration over raw joint targets into selection among pretrained behaviors; only the high-level policy is trained, and the frozen primitives provide both the search space and a natural regularizer against reward hacking.

What would settle it

Evaluate SkillBlender on a ninth task whose required subgoal lies outside each primitive's training range, for example a walking velocity command far beyond the low-level policy's training commands; if the high-level controller cannot compensate and task error stays above the success threshold while from-scratch PPO succeeds, the central claim of versatile blending collapses.

Watch

Extended reading notes

Core claim

SkillBlender's central discovery is that freezing four goal-conditioned primitive policies (walking, reaching, squatting, stepping) and training only a high-level controller to issue subgoals and per-joint blending weights produces more accurate and more physically feasible whole-body behaviors than learning each task from scratch or from a single low-level policy. The high-level controller's action is a softmax-weighted, elementwise combination of the primitives' target joint positions, and the softmax nonlinearity is claimed to be essential because a plain linear blend allows reward hacking. Across SkillBench's eight tasks and three humanoid embodiments, the paper reports that the blended policy either matches or beats all baselines on task error while exhibiting lower tilt and torque, and that ablations removing any component degrade performance.

Load-bearing premise

The load-bearing premise is that the four pretrained primitive policies stay stable and expressive when the high-level controller feeds them subgoals they never saw during training; the paper's own appendix concedes these policies can fail under out-of-distribution states such as unusual initial poses.

Editorial extensions

If this is right

  • New tasks need at most two reward terms, so task coverage scales as far as the skill library covers the required capabilities.
  • Blending beats sequential switching for humanoid whole-body tasks, because simultaneous activation lets walking and reaching act on different body parts at once.
  • Removing the softmax layer degrades feasibility metrics, so nonlinear per-joint weighting is doing real work, not incidental.
  • The same skill library transfers across three embodiments with only goal rescaling, which is evidence that the pretrained skills are reusable.
  • Skill selection can be automated by a foundation model, making the pipeline easier to deploy on new tasks.

Reading between the lines

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

  • If the pretrain-then-blend claim holds, the bottleneck shifts upstream to the skill library: a richer and more diverse set of primitives should extend the framework to bimanual, dexterous, and tool-use tasks without changing the learning machinery.
  • The benchmark's feasibility metrics may serve as a rough proxy for sim2real readiness, since policies with near-zero tilt and low torque are more likely to transfer than high-torque reward-hacking ones; testing that correlation is an open question.
  • The per-joint softmax weighting suggests a broader design principle: for high-dimensional whole-body control, constraining actions to weighted combinations of a few interpretable behaviors may be more effective than letting a monolithic policy explore freely, which could be tested on other robot platforms.
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

5 major / 5 minor

Summary. The paper introduces SkillBlender, a hierarchical reinforcement learning framework for humanoid whole-body loco-manipulation. The method first pretrains four goal-conditioned, task-agnostic low-level skills (Walking, Reaching, Squatting, Stepping) and then trains a high-level controller that outputs subgoals and per-joint softmax weights to blend the frozen low-level skills. The authors also introduce SkillBench, a parallel, cross-embodiment Isaac Gym benchmark with three humanoid embodiments, eight tasks, and accuracy plus feasibility metrics. Experiments compare SkillBlender against PPO, DreamerV3, a HumanoidBench-style baseline, Sequential HRL, and MCP on H1, and against PPO on G1 and H1-2, with ablations on skill subsets, the softmax weighting, and motion-tracking low-level policies, plus a real-world sanity check of primitive skills.

Significance. If the claims hold, SkillBlender's pretrain-then-blend paradigm is a worthwhile step toward reusable, modular humanoid skills that reduce per-task reward engineering. SkillBench itself is a useful contribution: it is parallel, cross-embodiment, includes complementary accuracy and feasibility metrics, and is planned for release. The paper also provides extensive training details, ablations, and a real-world deployment sanity check. However, the central claim of significantly outperforming all baselines in both accuracy and feasibility is not fully supported by the reported tables, the feasibility metrics lack statistical testing, the skill selector is manual in all experiments, and the robustness of frozen low-level skills under high-level-issued subgoals is not examined. These gaps do not invalidate the approach, but they require a major revision of the claims and additional analysis.

major comments (5)
  1. [Abstract and Section 5.2] The claim that SkillBlender 'significantly outperforms all baselines' in both accuracy and feasibility is not supported by Tables 1-3. For example, in ButtonPress (Table 1), MCP achieves a lower error (0.005±0.003 vs. 0.009±0.007) and better feasibility on every reported metric (Tilt 0.016 vs. 0.041, h 0.910 vs. 0.848, τ 13.9 vs. 16.8, P 19.2 vs. 20.3). In FarReach, PPO has the lowest error (0.016±0.008 vs. 0.021±0.012 for Ours) and MCP has better Tilt and h. MCP also shows better feasibility on several Medium and Hard tasks even when it fails the success threshold. The cross-embodiment results (Tables 5-10) compare only against PPO, not the full baseline set, so 'all baselines' is not established on G1 and H1-2. The comparative claims should be rewritten to reflect the actual pattern: SkillBlender is best on tasks where baselines fail (most Medium/Hard tasks) but is not uniformly superior on easy tasks or in feasibility.
  2. [Section 5.1] Feasibility metrics (Tilt, h, τ, P) are reported only as means across 20 rollouts, without standard deviations, error bars, or significance tests. This makes it impossible to assess whether differences such as τ=13.5 vs. 14.1 or Tilt=0.045 vs. 0.018 are meaningful. Given the abstract's explicit 'significantly outperforms' wording, the authors should either report per-metric variance or run paired significance tests across the 20 rollouts; otherwise the significance claims are unsupported.
  3. [Section 3.3 and Appendix I] The central pretrain-then-blend claim assumes that the frozen low-level skills remain competent on the subgoals and weight combinations produced by the high-level controller. The raw subgoals are only clamped (Section 3.3), not rescaled to the low-level training distribution, and the high-level controller is trained with PPO, so the distribution of issued subgoals is learned, not matched to pretraining. Appendix I explicitly concedes that primitive policies can fail under out-of-distribution states such as unusual initial poses. Because the low-level skills are frozen and the high-level controller only modulates their outputs, a broken low-level skill cannot be repaired. No experiment measures whether high-level-issued subgoals remain in the competent region, or whether blended actions degrade individual skill performance. This is a load-bearing gap for the claimed versatility and should be addressed with a subgoal-distribution analysis or a targeted stress test.
  4. [Section E.1] The skill selector is manual in all experiments reported in the paper; the GPT-4o example in Fig. 5 is illustrative only and is not evaluated as part of the benchmark. Thus the pipeline still requires human intervention to choose the relevant skill subset for each task. This significantly qualifies the 'versatile' and 'scalable' claims. The authors should either incorporate and evaluate the LLM-based selector, or clearly frame the current results as requiring a manually chosen skill library per task.
  5. [Section 3.2] The abstract's 'minimal task-specific reward engineering' claim is precise only for the high-level tasks. The low-level primitive skills are trained with dense rewards including 'task-relevant goal-matching rewards, regularization rewards, gait rewards, and other auxiliary objectives' (Section 3.2), and the exact reward terms are not documented anywhere in the main text or appendix. Thus the total reward-engineering effort is deferred to the pretraining stage, and the paper's phrasing could mislead. The authors should clarify that the one-or-two-term simplicity applies only to the high-level controller, and should disclose the low-level reward functions for reproducibility.
minor comments (5)
  1. [Section 3.3, Eq. (1)] The text states that each raw weight vector \~W^i_t is in [0,1]^d, but then applies a softmax; softmax is typically applied to unbounded logits. Please clarify whether the raw weights are indeed bounded or whether this is a typo for unbounded network outputs.
  2. [Section 4.4] The Tilt metric is described as 'the average root pitch and roll angles,' but it is unclear whether this is the mean absolute value or the signed mean. For a meaningful feasibility measure, absolute values are more natural; please specify.
  3. [Appendix D, Eqs. (4)-(11)] Several reward expressions use notations like '5e-4' and 'e-4' inconsistently (e.g., Eq. (7) has 'e−4' and '5e−'), and the meaning of '||·||' as Euclidean norm is not stated. A consistent notation would improve readability and reproducibility.
  4. [Section 5.3, Table 4] The ablation 'w/o Softmax' is evaluated on only three tasks; since the paper argues the softmax is a crucial anti-reward-hacking component, reporting it on additional tasks (especially hard ones like BoxTransfer) would strengthen the claim.
  5. [Section E.3] The heading contains a typo: 'Architechtures' should be 'Architectures'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: SkillBlender's pretrain-then-blend pipeline is a genuine compositional search, not an identity or fitted prediction.

full rationale

The paper's central derivation is hierarchical: task-agnostic low-level skills are pretrained with dense goal-conditioned rewards, and a separate high-level controller is trained with PPO to emit subgoals and per-joint weights that blend the frozen skills (Eqs. 1-2). The final action is a weighted sum of skill outputs, but the high-level controller must learn which subgoal sequences and weight combinations reduce the task-specific error; it is not simply inverting the low-level training objective. The task rewards in Appendix D are expressed through object and wrist distances, while the low-level skills are conditioned on qualitatively different goal spaces (velocity commands, wrist targets, root height, foot targets), so the mapping from task reward to high-level output is not definitionally forced. No fitted parameter is renamed as a prediction, and no uniqueness theorem or load-bearing self-citation is used; the self-citations (Refs. 10, 25, 26) support unrelated infrastructure or prior author work. The manual skill selection, author-chosen success thresholds, and author-designed benchmark are evaluation-validity concerns rather than circularity, and Appendix I's concession that primitive policies can fail under out-of-distribution states is a robustness limitation, not a circular reduction. The benchmark comparisons still reflect genuinely trained behavior, so the central claim is not equivalent to its inputs and no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hand-chosen, unauditable quantities: per-task reward coefficients, success thresholds, low-level dense reward weights, and the manual skill subset selection. The method introduces no new physical entities, but it does rely on a set of domain assumptions about simulation fidelity and PD control that are standard in the field yet unverified for this specific benchmark.

free parameters (4)
  • Task reward coefficients = 5e-4, e-4, 0.5e-4 per task term
    Each of the eight tasks in Appendix D uses hand-set coefficients before each reward term. These constants are not learned and directly affect learning behavior, so they count as free parameters.
  • Success thresholds = 0.01 to 1.5 m depending on task
    Hand-set thresholds determine which policies are counted as successful in Tables 1-3 and therefore influence the headline claim that SkillBlender solves all tasks.
  • Low-level dense reward weights = not reported
    Primitive skills are trained with dense rewards mixing goal, regularization, gait, and auxiliary terms, but the exact weights are not listed. This makes the amount of tuning for the supposedly task-agnostic skills unauditable.
  • Manually selected skill subset per task = 2-3 skills per task
    The high-level skill selector is manual in all experiments; the LLM example in Figure 5 is not used for the reported results. This is a human-chosen configuration that affects every task outcome.
assumptions (4)
  • domain assumption PD control with fixed gains converts target joint positions to torques, and this actuation model is sufficient for the reported behaviors.
    Invoked in Section 3.2 and used throughout the experiments, this is a standard robotics assumption but is load-bearing for the validity of the simulated results.
  • domain assumption Isaac Gym with the PhysX engine faithfully approximates humanoid dynamics for training and evaluation.
    All quantitative claims come from this simulator; the paper does not provide real-world validation of the high-level blended policies.
  • ad hoc to paper The four primitive skills form a sufficient basis for the eight benchmark tasks, and the manually chosen skill subsets are appropriate.
    Section 3.3 and Appendix E.1 rely on this premise. If a task needed a skill outside the library, the framework would fail, and the manual selection is not justified by any formal criterion.
  • ad hoc to paper The per-joint softmax in Eq. 1 is necessary to prevent reward hacking and to produce feasible motions.
    The softmax is introduced as a design choice and only supported by the ablation in Table 4. There is no derivation showing it is the unique or optimal normalizer for the blending operation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SkillBlender: Towards Versatile Humanoid Whole-Body Loco-Manipulation via Skill Blending." pith.science (2026). https://pith.science/paper/5H55TE7T

@misc{pith2026250609366,
  author       = {Pith},
  title        = {Pith review of: SkillBlender: Towards Versatile Humanoid Whole-Body Loco-Manipulation via Skill Blending},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5H55TE7T}},
  note         = {Machine review of arXiv:2506.09366}
}
read the original abstract

Humanoid robots hold significant potential in accomplishing daily tasks across diverse environments thanks to their flexibility and human-like morphology. Recent works have made significant progress in humanoid whole-body control and loco-manipulation leveraging optimal control or reinforcement learning. However, these methods require tedious task-specific tuning for each task to achieve satisfactory behaviors, limiting their versatility and scalability to diverse tasks in daily scenarios. To that end, we introduce SkillBlender, a novel hierarchical reinforcement learning framework for versatile humanoid loco-manipulation. SkillBlender first pretrains goal-conditioned task-agnostic primitive skills, and then dynamically blends these skills to accomplish complex loco-manipulation tasks with minimal task-specific reward engineering. We also introduce SkillBench, a parallel, cross-embodiment, and diverse simulated benchmark containing three embodiments, four primitive skills, and eight challenging loco-manipulation tasks, accompanied by a set of scientific evaluation metrics balancing accuracy and feasibility. Extensive simulated experiments show that our method significantly outperforms all baselines, while naturally regularizing behaviors to avoid reward hacking, resulting in more accurate and feasible movements for diverse loco-manipulation tasks in our daily scenarios. Our code and benchmark will be open-sourced to the community to facilitate future research. Project page: https://usc-gvl.github.io/SkillBlender-web/.

Figures

Figures reproduced from arXiv: 2506.09366 by the authors.

Figure 1
Figure 1. SkillBlender performs versatile autonomous humanoid loco-manipulation tasks within different embodiments and environments, given only one or two intuitive reward terms. Abstract Humanoid robots hold significant potential in accomplishing daily tasks across diverse environments thanks to their flexibility and human-like morphology. Recent works have made significant progress in humanoid whole-body control and loco￾ma… view at source ↗
Figure 2
Figure 2. Overview of SkillBlender. We first pretrain goal-conditioned primitive expert skills that are task-agnostic, reusable, and physically interpretable, and then reuse and blend these skills to achieve complex whole-body loco-manipulation tasks given only one or two task-specific reward terms. used in various daily tasks, and they also have physical goals like walking to one specific location, reaching a specific target… view at source ↗
Figure 3
Figure 3. Our SkillBench is a parallel, cross-embodiment, and diverse simulated benchmark contain￾ing three embodiments, four primitive skills, and eight loco-manipulation tasks. 4 SkillBench To facilitate standardized humanoid learning research, we propose a new benchmark SkillBench, which is parallel, cross-embodiment, and diverse. We implement SkillBench in NVIDIA Isaac Gym [34] with PhysX physics engine, benefiting from i… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between different methods. Our [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: An example of GPT-4o reasoning to perform skill selection on the [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on G1 and H1-2 embodiments. Our method produces more accurate and [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Visualization of whole-body per-joint weights at different stages of three different tasks. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Ego-centric visual observations, including RGB, depth (point cloud), and segmentation [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Demonstrations of our primitive skill sim2real deployment. We control the humanoid to [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Developing Combined Manipulation and Locomotion Skills with Interaction Representation and Skill Composition

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A simulated humanoid learns to grasp unseen objects and then stand up and walk while holding them, using a cubic-harmonics spatial representation and a finger-decoupling curriculum.

  2. Self-Adaptive Learning and Model Predictive Control for Tracking Unknown Dynamics with No Regret

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A self-adaptive MPC with multiple online-learned RFF predictors and Hedge-based selection achieves O(T^{3/4}) expected regret for tracking unknown, switching target dynamics.

Reference graph

Works this paper leans on

54 extracted references · 15 canonical work pages · cited by 2 Pith papers

  1. [1]

    Locomujoco: A comprehensive imitation learning benchmark for locomotion

    Firas Al-Hafez, Guoping Zhao, Jan Peters, and Davide Tateo. Locomujoco: A comprehensive imitation learning benchmark for locomotion. arXiv preprint arXiv:2311.02496, 2023

    arXiv 2023

  2. [2]

    The option-critic architecture

    Pierre-Luc Bacon, Jean Harb, and Doina Precup. The option-critic architecture. In Proceedings of the AAAI conference on artificial intelligence, volume 31, 2017

    2017

  3. [3]

    Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit

    Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin, and Jiangmiao Pang. Homie: Humanoid loco-manipulation with isomorphic exoskeleton cockpit. arXiv preprint arXiv:2502.13013, 2025

    arXiv 2025

  4. [4]

    Legs as manipulator: Pushing quadrupedal agility beyond locomotion

    Xuxin Cheng, Ashish Kumar, and Deepak Pathak. Legs as manipulator: Pushing quadrupedal agility beyond locomotion. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5106–5112. IEEE, 2023

    work page 2023

  5. [5]

    Expres- sive whole-body control for humanoid robots

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

    arXiv 2024

  6. [6]

    Bigym: A demo-driven mobile bi-manual manipulation benchmark

    Nikita Chernyadev, Nicholas Backshall, Xiao Ma, Yunfan Lu, Younggyo Seo, and Stephen James. Bigym: A demo-driven mobile bi-manual manipulation benchmark. arXiv preprint arXiv:2407.07788, 2024

    arXiv 2024

  7. [7]

    Adversarial motion priors make good substitutes for complex reward functions

    Alejandro Escontrela, Xue Bin Peng, Wenhao Yu, Tingnan Zhang, Atil Iscen, Ken Goldberg, and Pieter Abbeel. Adversarial motion priors make good substitutes for complex reward functions. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 25–32. IEEE, 2022

    work page 2022

  8. [8]

    Humanplus: Hu- manoid shadowing and imitation from humans

    Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Hu- manoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454, 2024

    arXiv 2024

Show all 54 references
  1. [9]

    Stochastic and robust mpc for bipedal locomotion: A comparative study on robustness and performance

    Ahmad Gazar, Majid Khadiv, Andrea Del Prete, and Ludovic Righetti. Stochastic and robust mpc for bipedal locomotion: A comparative study on robustness and performance. In 2020 IEEE-RAS 20th International Conference on Humanoid Robots (Humanoids) , pages 61–68. IEEE, 2021

    2020

  2. [10]

    Roboverse: Towards a unified plat- form, dataset and benchmark for scalable and generalizable robot learning

    Haoran Geng, Feishi Wang, Songlin Wei, Yuyang Li, Bangjun Wang, Boshi An, Charlie Tianyue Cheng, Haozhe Lou, Peihao Li, Yen-Jen Wang, et al. Roboverse: Towards a unified plat- form, dataset and benchmark for scalable and generalizable robot learning. arXiv preprint arXiv:2504....

    2025 arXiv

  3. [11]

    Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning

    Xinyang Gu, Yen-Jen Wang, Xiang Zhu, Chengming Shi, Yanjiang Guo, Yichen Liu, and Jianyu Chen. Advancing humanoid locomotion: Mastering challenging terrains with denoising world model learning. arXiv preprint arXiv:2408.14472, 2024

    2024 arXiv

  4. [12]

    Mastering diverse domains through world models

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023

    2023 arXiv

  5. [13]

    Hierarchical world models as visual whole-body humanoid controllers

    Nicklas Hansen, Jyothir SV , Vlad Sobal, Yann LeCun, Xiaolong Wang, and Hao Su. Hierarchical world models as visual whole-body humanoid controllers. arXiv preprint arXiv:2405.18418, 2024

    2024 arXiv

  6. [14]

    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. arXiv preprint arXiv:2406.08858, 2024

    2024 arXiv

  7. [15]

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

    Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation. arXiv preprint arXiv:2403.04436, 2024

    2024 arXiv

  8. [16]

    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. arXiv preprint arXiv:2410.21229, 2024

    arXiv

  9. [17]

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

    Tairan He, Jiawei Gao, Wenli Xiao, Yuanhang Zhang, Zi Wang, Jiashun Wang, Zhengyi Luo, Guanqi He, Nikhil Sobanbab, Chaoyi Pan, et al. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills. arXiv preprint arXiv:2502.01143, 2025

    2025 arXiv

  10. [18]

    Learning getting-up policies for real-world humanoid robots

    Xialin He, Runpei Dong, Zixuan Chen, and Saurabh Gupta. Learning getting-up policies for real-world humanoid robots. arXiv preprint arXiv:2502.12152, 2025

    2025 arXiv

  11. [19]

    Learning and transfer of modulated locomotor controllers.arXiv preprint arXiv:1610.05182, 2016

    Nicolas Heess, Greg Wayne, Yuval Tassa, Timothy Lillicrap, Martin Riedmiller, and David Sil- ver. Learning and transfer of modulated locomotor controllers.arXiv preprint arXiv:1610.05182, 2016

    2016 arXiv

  12. [20]

    Learning humanoid standing-up control across diverse postures

    Tao Huang, Junli Ren, Huayi Wang, Zirui Wang, Qingwei Ben, Muning Wen, Xiao Chen, Jianan Li, and Jiangmiao Pang. Learning humanoid standing-up control across diverse postures. arXiv preprint arXiv:2502.08378, 2025

    2025 arXiv

  13. [21]

    Modeling, stability and control of biped robots—a general framework

    Yildirim Hurmuzlu, Frank Génot, and Bernard Brogliato. Modeling, stability and control of biped robots—a general framework. Automatica, 40(10):1647–1664, 2004

    2004

  14. [22]

    Exbody2: Advanced expressive humanoid whole-body control

    Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiao- long Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196, 2024

    2024 arXiv

  15. [23]

    Hierarchical reinforcement learning for precise soccer shooting skills using a quadrupedal robot

    Yandong Ji, Zhongyu Li, Yinan Sun, Xue Bin Peng, Sergey Levine, Glen Berseth, and Koushil Sreenath. Hierarchical reinforcement learning for precise soccer shooting skills using a quadrupedal robot. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IR...

    2022

  16. [24]

    Tailoring solution accuracy for fast whole-body model predictive control of legged robots

    Charles Khazoom, Seungwoo Hong, Matthew Chignoli, Elijah Stanger-Jones, and Sangbae Kim. Tailoring solution accuracy for fast whole-body model predictive control of legged robots. IEEE Robotics and Automation Letters, 2024

    2024

  17. [25]

    Openfmnav: Towards open-set zero-shot object navigation via vision-language foundation models

    Yuxuan Kuang, Hai Lin, and Meng Jiang. Openfmnav: Towards open-set zero-shot object navigation via vision-language foundation models. arXiv preprint arXiv:2402.10670, 2024

    2024 arXiv

  18. [26]

    Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation

    Yuxuan Kuang, Junjie Ye, Haoran Geng, Jiageng Mao, Congyue Deng, Leonidas Guibas, He Wang, and Yue Wang. Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation. arXiv preprint arXiv:2407.04689, 2024

    2024 arXiv

  19. [27]

    Cascaded compositional residual learning for complex interactive behaviors

    K Niranjan Kumar, Irfan Essa, and Sehoon Ha. Cascaded compositional residual learning for complex interactive behaviors. IEEE Robotics and Automation Letters, 8(8):4601–4608, 2023. 11

    2023

  20. [28]

    Learning quadrupedal locomotion over challenging terrain

    Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5(47):eabc5986, 2020

    2020

  21. [29]

    Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators

    Chengshu Li, Fei Xia, Roberto Martin-Martin, and Silvio Savarese. Hrl4in: Hierarchical reinforcement learning for interactive navigation with mobile manipulators. In Conference on Robot Learning, pages 603–616. PMLR, 2020

    2020

  22. [30]

    Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control, 2025

    Jialong Li, Xuxin Cheng, Tianshu Huang, Shiqi Yang, Ri-Zhao Qiu, and Xiaolong Wang. Amo: Adaptive motion optimization for hyper-dexterous humanoid whole-body control, 2025. URL https://arxiv.org/abs/2505.03738

    2025 arXiv

  23. [31]

    Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking

    Yun Liu, Bowen Yang, Licheng Zhong, He Wang, and Li Yi. Mimicking-bench: A benchmark for generalizable humanoid-scene interaction learning via human mimicking. arXiv preprint arXiv:2412.17730, 2024

    2024 arXiv

  24. [32]

    Perpetual humanoid control for real-time simulated avatars

    Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895–10904, 2023

    2023

  25. [33]

    Universal humanoid motion representations for physics-based control

    Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. arXiv preprint arXiv:2310.04582, 2023

    2023 arXiv

  26. [34]

    Isaac gym: High performance gpu-based physics simulation for robot learning, 2021

    Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, and Gavriel State. Isaac gym: High performance gpu-based physics simulation for robot learning, 2021

    2021

  27. [35]

    Learning robust perceptive locomotion for quadrupedal robots in the wild

    Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Science robotics, 7(62):eabk2822, 2022

    2022

  28. [36]

    Mcp: Learning composable hierarchical control with multiplicative compositional policies

    Xue Bin Peng, Michael Chang, Grace Zhang, Pieter Abbeel, and Sergey Levine. Mcp: Learning composable hierarchical control with multiplicative compositional policies. Advances in neural information processing systems, 32, 2019

    2019

  29. [37]

    Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters

    Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Transactions On Graphics (TOG), 41(4):1–17, 2022

    2022

  30. [38]

    Humanoid locomotion as next token prediction

    Ilija Radosavovic, Bike Zhang, Baifeng Shi, Jathushan Rajasegaran, Sarthak Kamat, Trevor Darrell, Koushil Sreenath, and Jitendra Malik. Humanoid locomotion as next token prediction. arXiv preprint arXiv:2402.19469, 2024

    2024 arXiv

  31. [39]

    Human growth and development-a matter of principles

    Novella J Ruffin. Human growth and development-a matter of principles. 2009

    2009

  32. [40]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  33. [41]

    Hu- manoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation

    Carmelo Sferrazza, Dun-Ming Huang, Xingyu Lin, Youngwoon Lee, and Pieter Abbeel. Hu- manoidbench: Simulated humanoid benchmark for whole-body locomotion and manipulation. arXiv preprint arXiv:2403.10506, 2024

    2024 arXiv

  34. [42]

    Neural state machine for character- scene interactions

    Sebastian Starke, He Zhang, Taku Komura, and Jun Saito. Neural state machine for character- scene interactions. ACM Transactions on Graphics, 38(6):178, 2019

    2019

  35. [43]

    Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning

    Richard S Sutton, Doina Precup, and Satinder Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial intelligence, 112(1-2): 181–211, 1999

    1999

  36. [44]

    Deepmind control suite

    Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Budden, Abbas Abdolmaleki, Josh Merel, Andrew Lefrancq, et al. Deepmind control suite. arXiv preprint arXiv:1801.00690, 2018. 12

    2018 arXiv

  37. [45]

    Revisiting reward design and evaluation for robust humanoid standing and walking

    Bart van Marum, Aayam Shrestha, Helei Duan, Pranay Dugar, Jeremy Dao, and Alan Fern. Revisiting reward design and evaluation for robust humanoid standing and walking. arXiv preprint arXiv:2404.19173, 2024

    2024 arXiv

  38. [46]

    Unicon: Universal neural controller for physics-based character motion

    Tingwu Wang, Yunrong Guo, Maria Shugrina, and Sanja Fidler. Unicon: Universal neural controller for physics-based character motion. arXiv preprint arXiv:2011.15119, 2020

    2011 arXiv

  39. [47]

    Skillmimic: Learning reusable basketball skills from demonstrations

    Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, et al. Skillmimic: Learning reusable basketball skills from demonstrations. arXiv preprint arXiv:2408.15270, 2024

    2024 arXiv

  40. [48]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  41. [49]

    Multi-expert learning of adaptive legged locomotion

    C Yang, K Yuan, Q Zhu, W Yu, and Z Li. Multi-expert learning of adaptive legged locomotion. sci, 2020

    2020

  42. [50]

    Twist: Teleoperated whole-body imitation system

    Yanjie Ze, Zixuan Chen, Joà G, 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

  43. [51]

    Wococo: Learning whole-body humanoid control with sequential contacts

    Chong Zhang, Wenli Xiao, Tairan He, and Guanya Shi. Wococo: Learning whole-body humanoid control with sequential contacts. arXiv preprint arXiv:2406.06005, 2024

    2024 arXiv

  44. [52]

    Gamma: Graspability-aware mobile manipulation policy learning based on online grasping pose fusion

    Jiazhao Zhang, Nandiraju Gireesh, Jilong Wang, Xiaomeng Fang, Chaoyi Xu, Weiguang Chen, Liu Dai, and He Wang. Gamma: Graspability-aware mobile manipulation policy learning based on online grasping pose fusion. In 2024 IEEE International Conference on Robotics and Automation (I...

    2024

  45. [53]

    Embrace collisions: Humanoid shadowing for deployable contact-agnostics motions

    Ziwen Zhuang and Hang Zhao. Embrace collisions: Humanoid shadowing for deployable contact-agnostics motions. arXiv preprint arXiv:2502.01465, 2025

    2025 arXiv

  46. [54]

    Humanoid parkour learning

    Ziwen Zhuang, Shenzhe Yao, and Hang Zhao. Humanoid parkour learning. arXiv preprint arXiv:2406.10759, 2024. 13 A Observation Space For state-based policies, the observation space for the actor (goalgt and state st) comprises 3d+6+ N dimensions (d is the robot’s DoF), in which ...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.