Pith. sign in

REVIEW 3 major objections 26 references

Hard humanoid motions stay unsolved under the usual recipe because the recipe itself limits what the controller can learn.

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 · grok-4.5

2026-07-11 12:45 UTC pith:SSY6JWMR

load-bearing objection Useful systems paper: residual long-tail failures as recipe/capability mismatch, with a compact dynamic/balance expert pipeline and better high-coverage metrics—solid sim evidence, incomplete residual isolation. the 3 major comments →

arxiv 2607.04837 v1 pith:SSY6JWMR submitted 2026-07-06 cs.RO

Athena-WBC: Capability-Aligned Policy Experts for Long-Tail Humanoid Whole-Body Control

Yuan Jiang , Ningyuan Zhang , Xicun Yang , Shidi Li , Yuzhi Jiang , Zhiyi Rong , Shuaikang Ma , Chuanzheng Li
show 1 more author
Jie Chen
This is my paper
classification cs.RO
keywords humanoid whole-body controlmotion trackinglong-tail learningcapability bottleneckteacher-student distillationDAggergravity curriculumpolicy regularization
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.

Strong humanoid whole-body trackers still leave a residual set of feasible training motions unsolved, especially high-dynamic transitions and balance-critical poses. The paper argues this is not only a data-allocation problem: even when those clips are oversampled or trained in isolation, the default reward recipe—effort and smoothness penalties under full gravity—can suppress the aggressive yet feasible actions or early survivability the motions need. Athena-WBC trains a small bank of capability-aligned privileged experts: dynamic experts drop conservative effort and temporal-control rewards while keeping physical limits and imposing smoothness via an auxiliary policy loss; balance experts start under reduced gravity so early rollouts survive long enough to learn. Those teachers are then routed per motion, distilled into one deployable student, and fine-tuned with RL. On a full-size humanoid, the pipeline recovers more of the training long tail and improves held-out tracking versus a strong SONIC-recipe baseline, with only a few experts.

Core claim

In strong humanoid whole-body control baselines, residual feasible training clips remain unsolved even under targeted training because of a capability bottleneck: a mismatch between motion demands and the effective control regime induced by the default acquisition recipe. Capability-aligned experts—dynamic experts that remove effort and temporal-control penalties while retaining physical constraints and balance experts that use a gravity curriculum—plus motion-routed distillation and RL fine-tuning recover more of that long tail and improve held-out tracking than reallocating data alone under the same recipe.

What carries the argument

Capability-aligned policy experts: dynamic experts train with tracking plus physical-constraint rewards only, plus Grad-CAPS auxiliary smoothness on the policy mean; balance experts train under a gravity-scale curriculum; both residual-set teachers are then motion-routed for DAgger distillation into one student that is RL-finetuned under deployable observations.

Load-bearing premise

That residual failures after data-only interventions mainly reflect a recipe-induced capability mismatch rather than imperfect retargeted references, near-limit physical infeasibility, embodiment limits, or under-optimized training budgets.

What would settle it

Train the same residual clips to saturation under the default recipe with matched budget and cleaner references; if those clips then succeed at high rate without capability changes, the bottleneck claim fails. Conversely, if dynamic and balance recipe changes still do not recover a large share of the residual set, the expert design is insufficient.

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

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

3 major / 0 minor

Summary. Athena-WBC argues that residual training-set failures in strong humanoid whole-body motion trackers are not only data-allocation problems but capability bottlenecks induced by the default acquisition recipe. On a full-size 80 kg humanoid, the authors reimplement a SONIC-style baseline, mine residual clips from a general privileged teacher, and train two capability-aligned experts on the same residual set: a dynamic expert that keeps tracking and physical-constraint rewards while removing effort/temporal penalties and regularizing with Grad-CAPS, and a balance expert trained with a gravity curriculum. Teachers are motion-routed by rollout success, distilled with DAgger into one deployable student, and RL-finetuned. Relative to SONIC-Base, the final policy improves held-out AMASS/Omni SR, TIS, MPJPE, and MPJPE-W, with ablations on no-smoothness, CAPS vs Grad-CAPS, single- vs multi-teacher distillation, and teacher observations, plus new diagnostics (STC/TIS, MPJPE-W).

Significance. If the residual-failure diagnosis holds, the paper usefully reframes long-tail humanoid WBC: changing reward/curriculum capability can matter more than only resampling or partitioning motions, and a compact two-expert bank can recover complementary high-dynamic and balance regimes before distillation. The empirical package is a real strength for a systems paper—multi-seed tables, smoothness-placement ablations with spectra/jitter, qualitative case studies, and threshold-/salience-aware metrics that are well motivated in the high-coverage regime. The contribution is incremental relative to SONIC/OmniH2O/expert-distillation recipes, but the capability-alignment framing and evaluation tools are of clear interest to the humanoid control community.

major comments (3)
  1. The central claim that residual failures are primarily recipe-induced capability bottlenecks (Secs. 1, 3.2–3.3; Eqs. 6–8) is only partially isolated from reference quality and physical feasibility. Rgen is defined by cSR < 0.8 on the general teacher; Fig. 2 and the Limitations section already admit data artifacts, near-limit motions, and incomplete teacher acquisition, yet Table 4 reports recovery on automatically constructed Ddynamic/Dbalance manifests (App. A.4) rather than on a verified recoverable residual subset with per-clip feasibility/reference filters. Please quantify what fraction of Rgen is recovered by no-smoothness/experts versus remains unsolved for reference/embodiment reasons, and report long-tail metrics on that filtered residual set.
  2. The abstract and introduction claim residual clips remain unsolved 'even under targeted training,' but the main experimental tables do not include a controlled data-only intervention baseline on the same residual set (e.g., oversampling or training only on Rgen under the SONIC-Base recipe with matched budget). Without that comparison, Table 4’s expert gains show that less conservative recipes track hard clips better, but do not fully establish that exposure alone is insufficient. Add this ablation, or narrow the claim to the evidence actually reported.
  3. All quantitative results are simulation-only on a proprietary, unreleased platform (Sec. 5.1, Limitations). The paper correctly scopes claims to recipe-level comparison on one embodiment, but for a journal contribution in humanoid WBC the deployability claim (action-rate recovery, RL fine-tuning as deployment stage) needs at least limited real-robot quantitative tracking/smoothness results, or a clearer demotion of hardware-readiness claims until such evidence exists.

Circularity Check

0 steps flagged

No circular derivation: empirical systems claims measured by independent rollouts, not tautologies of fitted inputs.

full rationale

Athena-WBC is a methods/systems paper whose load-bearing claims are empirical: residual training-set failures under a SONIC-recipe baseline, recovery after changing acquisition recipes (remove effort/temporal reward terms; gravity curriculum), and held-out gains after routed DAgger distillation plus RL fine-tuning. Residual sets R(π0)/Rgen (Eqs. 6–8) are defined by measured rollout success of a trained general teacher, not by the expert objectives or final student metrics; experts are then trained with altered rewards/curricula and evaluated with fixed thresholds, STC/TIS, MPJPE, and MPJPE-W on training and held-out splits. That experimental dependence on a chosen ρ_fail and on the authors’ reimplementation is ordinary methodology, not a reduction of the claimed result to its inputs by construction. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from overlapping authors, and no ansatz smuggled in as a first-principles derivation. Concerns about residual-set confounds (reference quality, physical saturation) affect causal interpretation of the capability-bottleneck claim, not circularity of the derivation chain. Score 0 with empty steps is therefore the correct outcome.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 4 invented entities

The central claim rests on standard RL/control assumptions plus several paper-specific operational choices: residual clips with success below a hand-chosen threshold are treated as capability bottlenecks; effort/temporal reward terms are treated as the main suppressors of high-dynamic behavior; gravity scale is treated as a valid early-survivability curriculum variable; and simulator success under domain randomization is treated as a proxy for recoverable tracking skill. Free parameters include failure thresholds, curriculum ranges, regularization weights, and sampler constants. Invented entities are mostly methodological constructs (capability bottleneck framing, expert bank, and new evaluation metrics), not physical particles.

free parameters (6)
  • residual failure threshold ρ_fail
    Clips with cSR < 0.8 enter the residual set used to train capability experts; this threshold directly defines the long-tail problem the method claims to solve.
  • Grad-CAPS / CAPS regularization weight λ_reg
    Controls the tracking–smoothness trade-off after removing reward-level effort/temporal penalties; chosen experimentally and load-bearing for deployability claims.
  • gravity curriculum range α_e ∈ [α_min, 1]
    Determines early-training survivability for balance experts; the paper does not show uniqueness of this schedule.
  • adaptive sampling mixture ρ, EMA β_ema, violation bonus c_viol, difficulty cap D_max
    These sampler constants allocate training budget and are described as ‘small but critical’ differences from prior samplers; they affect which residual clips are learned.
  • default success thresholds δz=0.20 m, δyaw=0.50 rad, δkb=0.50 m
    Define SR and anchor STC/TIS; rankings can change with tolerance, so the reported gains partly depend on this evaluation convention.
  • training budget 40k iterations / 16384 envs and RL fine-tuning schedule
    All non-finetuned comparisons use this budget; residual unsolved clips could partly reflect optimization limits rather than pure capability mismatch.
axioms (6)
  • ad hoc to paper A residual clip that remains unsolved after data-only interventions but becomes solvable after changing reward/curriculum is a capability bottleneck relative to embodiment, simulator, protocol, and budget.
    Operational definition in Sec. 3.3; central to the paper’s reframing of long-tail failure.
  • domain assumption Physical-constraint penalties encode feasibility, while effort and temporal-control penalties encode conservative control preferences that can suppress feasible high-dynamic actions.
    Reward decomposition in Sec. 3.1 and dynamic-expert design in Sec. 4.1; supported by no-smoothness ablations but still a modeling assumption.
  • domain assumption Reduced-gravity continuation improves early survivability of balance-critical motions without changing the final evaluation dynamics.
    Balance-expert curriculum in Sec. 4.1; standard continuation-learning idea applied to gravity scale.
  • domain assumption Privileged teacher actions provide useful supervision for a deployable student via DAgger, and post-distillation RL fine-tuning improves closed-loop generalization.
    Standard teacher-student WBC assumption used throughout Sec. 4.3–4.4 and related work.
  • domain assumption Simulator rollouts under domain randomization with fixed success predicates are a valid proxy for recoverable tracking performance on the target embodiment.
    All quantitative claims are sim-based; real-robot quantitative validation is deferred (Limitations).
  • standard math PPO optimization, GAE, and standard continuous-control RL math are valid for the reported training procedure.
    Background optimization machinery; not re-derived.
invented entities (4)
  • capability bottleneck (recipe-induced effective control regime) no independent evidence
    purpose: Explain residual training-set failures that persist under targeted exposure.
    Core conceptual construct of the paper; operationalized via residual mining and recipe changes, but not independently measured outside this pipeline.
  • Athena-WBC capability-aligned expert bank (dynamic + balance teachers with motion routing) no independent evidence
    purpose: Acquire complementary hard-motion behaviors and compress them into one student.
    Methodological system introduced by the paper; evidence is internal experimental comparison.
  • Success–Tolerance Curve (STC) and Threshold-Integrated Success (TIS) no independent evidence
    purpose: Expose threshold sensitivity of high-coverage WBC success rates.
    New evaluation constructs defined in Sec. 5.2 and App. A.3; useful diagnostics but paper-defined.
  • Motion-Salience Weighted MPJPE (MPJPE-W) no independent evidence
    purpose: Emphasize tracking error on body parts that define each reference motion.
    New diagnostic metric based on reference local motion lengths; not an external physical quantity.

pith-pipeline@v1.1.0-grok45 · 25944 in / 4467 out tokens · 37150 ms · 2026-07-11T12:45:05.004836+00:00 · methodology

0 comments
read the original abstract

Large-scale humanoid motion-tracking controllers are commonly improved by reallocating training effort: difficult motions are sampled more often, isolated into smaller subsets, or assigned to specialized experts. We show that this view is incomplete. In strong whole-body-control baselines, a residual set of feasible training clips remains unsolved even under targeted training, especially for high-dynamic transitions and balance-critical motions. These failures arise not only from insufficient exposure, but from a mismatch between the motion demands and the effective capability induced by the default training recipe. We propose Athena-WBC, a compact teacher-student pipeline with capability-aligned policy experts for long-tail humanoid whole-body control. Dynamic experts use a tracking-focused, constraint-aware objective that removes conservative effort and temporal-control penalties while preserving physical feasibility constraints; balance experts use a gravity curriculum to improve early-training survivability. The resulting privileged teachers are motion-routed for DAgger distillation and then compressed into a single controller with deployable observations followed by RL fine-tuning. Experiments on a full-size humanoid show improved recovery of training-set long-tail motions and better held-out tracking than a strong SONIC-recipe baseline, using only a small number of experts.

Figures

Figures reproduced from arXiv: 2607.04837 by Chuanzheng Li, Jie Chen, Ningyuan Zhang, Shidi Li, Shuaikang Ma, Xicun Yang, Yuan Jiang, Yuzhi Jiang, Zhiyi Rong.

Figure 1
Figure 1. Figure 1: Overview of Athena-WBC. A general privileged teacher is trained on the full motion set. Residual failures are mined and used to train dynamic and balance experts in parallel. The frozen teachers are then routed per motion, distilled into a single student, and finetuned with RL. experts [11]. These strategies can improve coverage, but they primarily change which data a policy sees. They do not necessarily c… view at source ↗
Figure 2
Figure 2. Figure 2: Training-set long-tail failures in SONIC baseline. The released checkpoint leaves a small but [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Adaptive motion sampling. Rollout tracking errors are converted into smoothed temporal-bin [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative case study: crouch-and-walk-forward motion. The upper strip shows the SONIC [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative case study: high-kick motion with a large single-leg swing. The upper strip shows [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative case studies. In each case, the upper strip shows SONIC-Base and the lower strip [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MPJPE-W and STC as complementary evaluation tools for tracking WBC. Left: Motion [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Key-body-tolerance Success–Tolerance Curves on the two held-out evaluation sets. Unlike [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Clip-level high-frequency action-energy comparison against SONIC-Base. Each point is one [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Mean normalized action power spectral density under different smoothness placements. The [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Representative successful high-dynamic backward-running clip. The rows show an oscillatory [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] 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

26 extracted references · 1 canonical work pages

  1. [1]

    SONIC: Supersizing motion tracking for natural humanoid whole-body control, 2025

    Zhengyi Luo, Ye Yuan, Tingwu Wang, Chenran Li, Fernando Castañeda, Sirui Chen, Zi-Ang Cao, Jiefeng Li, David Minor, Qingwei Ben, Jinhyung Park, David Sami, Zi Wang, Xingye Da, Runyu Ding, Cyrus Hogg, Lina Song, Edy Lim, Eugene Jeong, Tairan He, Haoru Xue, Wenli Xiao, Simon Yuen, Jan Kautz, Yan Chang, Umar Iqbal, Linxi Jim Fan, and Yuke Zhu. SONIC: Supersi...

    Pith/arXiv arXiv 2025

  2. [2]

    Perpetual humanoid control for real-time simulated avatars

    Zhengyi Luo, Jinkun Cao, Alexander Winkler, Kris Kitani, and Weipeng Xu. Perpetual humanoid control for real-time simulated avatars. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10861–10870, 2023. doi: 10.1109/ICCV51070.2023.01000

    doi:10.1109/iccv51070.2023.01000 2023

  3. [3]

    Kitani, Changliu Liu, and Guanya Shi

    Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris M. Kitani, Changliu Liu, and Guanya Shi. OmniH2O: Universal and dexterous human-to-humanoid whole- body teleoperation and learning. In Pulkit Agrawal, Oliver Kroemer, and Wolfram Burgard, editors, Proceedings of The 8th Conference on Robot Learning, volume 270 ofProceedings of ...

    2025

  4. [4]

    EGM: Efficiently learning general motion tracking policy for high dynamic humanoid whole-body control, 2025

    Chao Yang, Yingkai Sun, Peng Ye, Xin Chen, Chong Yu, and Tao Chen. EGM: Efficiently learning general motion tracking policy for high dynamic humanoid whole-body control, 2025. URL https://arxiv.org/abs/2512.19043

    arXiv 2025

  5. [5]

    MOSAIC: Bridging the Sim-to-Real gap in generalist humanoid motion tracking and teleoperation with rapid residual adaptation, 2026

    Zhenguo Sun, Bo-Sheng Huang, Yibo Peng, Xukun Li, Jingyu Ma, Yu Sun, Zhe Li, Haojun Jiang, Biao Gao, Zhenshan Bing, Xinlong Wang, and Alois Knoll. MOSAIC: Bridging the Sim-to-Real gap in generalist humanoid motion tracking and teleoperation with rapid residual adaptation, 2026. URLhttps://arxiv.org/abs/2602.08594. 21

    arXiv 2026

  6. [6]

    RobotDancing: Residual-action reinforcement learning enables robust long-horizon humanoid motion tracking, 2025

    Zhenguo Sun, Yibo Peng, Yuan Meng, Xukun Li, Bo-Sheng Huang, Zhenshan Bing, Xinlong Wang, and Alois Knoll. RobotDancing: Residual-action reinforcement learning enables robust long-horizon humanoid motion tracking, 2025. URLhttps://arxiv.org/abs/2509.20717

    arXiv 2025

  7. [7]

    M3imic: Learning a versatile whole-body controller for multimodal motion mimicking, 2026

    Zuxing Lu, Ziang Zheng, Yao Lyu, Jingyu Liu, Feihong Zhang, Song Lu, Xin Yuan, Changyin Sun, Xingxing Zuo, and Shengbo Eben Li. M3imic: Learning a versatile whole-body controller for multimodal motion mimicking, 2026. URLhttps://arxiv.org/abs/2606.04829

    Pith/arXiv arXiv 2026

  8. [8]

    Stubborn: A streamlined and unified reinforcement learning framework for robust motion tracking and fall recovery for humanoids,

    Xiao Ren, Yuhui Yang, Zongbiao Weng, Zhijie Liu, and He Kong. Stubborn: A streamlined and unified reinforcement learning framework for robust motion tracking and fall recovery for humanoids,

  9. [9]

    URLhttps://arxiv.org/abs/2606.12814

    Pith/arXiv arXiv

  10. [10]

    Agility meets stability: Versatile humanoid control with heterogeneous data, 2025

    Yixuan Pan, Ruoyi Qiao, Li Chen, Kashyap Chitta, Liang Pan, Haoguang Mai, Qingwen Bu, Hao Zhao, Cunyuan Zheng, Ping Luo, and Hongyang Li. Agility meets stability: Versatile humanoid control with heterogeneous data, 2025. URLhttps://arxiv.org/abs/2511.17373

    arXiv 2025

  11. [11]

    From experts to a generalist: Toward general whole-body control for humanoid robots,

    Yuxuan Wang, Ming Yang, Weishuai Zeng, Yu Zhang, Xinrun Xu, Haobin Jiang, Ziluo Ding, and Zongqing Lu. From experts to a generalist: Toward general whole-body control for humanoid robots,

  12. [12]

    URLhttps://arxiv.org/abs/2506.12779

    Pith/arXiv arXiv

  13. [13]

    Humanoid-GPT: Scaling data and structure for zero-shot motion tracking, 2026

    Zekun Qi, Xuchuan Chen, Dairu Liu, Chenghuai Lin, Yunrui Lian, Sikai Liang, Zhikai Zhang, Yu Guan, Jilong Wang, Wenyao Zhang, Xinqiang Yu, He Wang, and Li Yi. Humanoid-GPT: Scaling data and structure for zero-shot motion tracking, 2026. URL https://arxiv.org/abs/2606. 03985

    2026

  14. [14]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Stephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Geoffrey Gordon, David Dunson, and Miroslav Dudík, editors,Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 ofProceedings of Machine Learning Research, pag...

    2011

  15. [15]

    GMT: General motion tracking for humanoid whole-body control, 2025

    Zixuan Chen, Mazeyu Ji, Xuxin Cheng, Xuanbin Peng, Xue Bin Peng, and Xiaolong Wang. GMT: General motion tracking for humanoid whole-body control, 2025. URL https://arxiv.org/abs/ 2506.14770

    Pith/arXiv arXiv 2025

  16. [16]

    Estrada, Francesco Iacobelli, Twan Koolen, Alexander Lambert, Erica Lin, M

    Jean Pierre Sleiman, He Li, Alphonsus Adu-Bredu, Robin Deits, Arun Kumar, Kevin Bergamin, Mohak Bhardwaj, Scott Biddlestone, Nicola Burger, Matthew A. Estrada, Francesco Iacobelli, Twan Koolen, Alexander Lambert, Erica Lin, M. Eva Mungai, Zach Nobles, Shane Rozen-Levy, Yuyao Shi, Jiashun Wang, Jakob Welner, Fangzhou Yu, Mike Zhang, Alfred Rizzi, Jessica H...

    arXiv 2026

  17. [17]

    OmniTrack: General motion tracking via physics-consistent reference, 2026

    Yuhan Li, Peiyuan Zhi, Yunshen Wang, Tengyu Liu, Sixu Yan, Wenyu Liu, Xinggang Wang, Baoxiong Jia, and Siyuan Huang. OmniTrack: General motion tracking via physics-consistent reference, 2026. URLhttps://arxiv.org/abs/2602.23832

    arXiv 2026

  18. [18]

    Regularizing action policies for smooth control with reinforcement learning

    Siddharth Mysore, Bassel Mabsout, Renato Mancuso, and Kate Saenko. Regularizing action policies for smooth control with reinforcement learning. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1810–1816, 2021. doi: 10.1109/ICRA48506.2021.9561138

    doi:10.1109/icra48506.2021.9561138 2021

  19. [19]

    Lee, Hoang-Giang Cao, Cong-Tinh Dao, Yu-Cheng Chen, and I-Chen Wu

    I. Lee, Hoang-Giang Cao, Cong-Tinh Dao, Yu-Cheng Chen, and I-Chen Wu. Gradient-based regularization for action smoothness in robotic control with reinforcement learning. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 603–610, 2024. doi: 10.1109/IROS58592.2024.10801464

    doi:10.1109/iros58592.2024.10801464 2024

  20. [20]

    Regularization matters in policy optimization – an empirical study on continuous control

    Zhuang Liu, Xuanlin Li, Bingyi Kang, and Trevor Darrell. Regularization matters in policy optimization – an empirical study on continuous control. InInternational Conference on Learning Representations, 2021. URLhttps://openreview.net/forum?id=yr1mzrH3IC. 22

    2021

  21. [21]

    AWAC: Accelerating online reinforcement learning with offline datasets, 2020

    Ashvin Nair, Abhishek Gupta, Murtaza Dalal, and Sergey Levine. AWAC: Accelerating online reinforcement learning with offline datasets, 2020. URLhttps://arxiv.org/abs/2006.09359

    Pith/arXiv arXiv 2020

  22. [22]

    Residual force control for agile human behavior imitation and extended motion synthesis

    Ye Yuan and Kris Kitani. Residual force control for agile human behavior imitation and extended motion synthesis. InAdvances in Neural Information Processing Sys- tems, volume 33, 2020. URL https://proceedings.neurips.cc/paper/2020/hash/ f76a89f0cb91bc419542ce9fa43902dc-Abstract.html

    2020

  23. [23]

    Learning motion skills with adaptive assistive curriculum force in humanoid robots, 2025

    Zhanxiang Cao, Yang Zhang, Buqing Nie, Huangxuan Lin, Haoyang Li, and Yue Gao. Learning motion skills with adaptive assistive curriculum force in humanoid robots, 2025. URL https: //arxiv.org/abs/2506.23125

    Pith/arXiv arXiv 2025

  24. [24]

    Parkour in the wild: Learning a general and extensible agile locomotion policy using multi-expert distillation and RL fine-tuning

    Nikita Rudin, Junzhe He, Joshua Aurand, and Marco Hutter. Parkour in the wild: Learning a general and extensible agile locomotion policy using multi-expert distillation and RL fine-tuning. The International Journal of Robotics Research, 2026. doi: 10.1177/02783649261455067. URL https://doi.org/10.1177/02783649261455067. OnlineFirst

    work page doi:10.1177/02783649261455067 2026

  25. [25]

    Troje, Gerard Pons-Moll, and Michael J

    Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5441–5450, 2019. doi: 10.1109/ICCV .2019.00554

    doi:10.1109/iccv 2019

  26. [26]

    BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis

    Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. BEAT: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. InComputer Vision – ECCV 2022, pages 612–630, 2022. doi: 10.1007/ 978-3-031-20071-7_36. 23 A Appendix A.1 Training Hyperparameters Table 7 lists the PP...

    2022