Pith. sign in

REVIEW 4 major objections 7 minor 23 references

Two humanoid roller-skating gaits can be learned from retargeted motion capture using adversarial motion priors and a passive-wheel simulation model.

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

2026-07-14 09:02 UTC pith:3MSREZFX

load-bearing objection Solid systems demo of two passive-skate humanoid gaits with AMP and a practical wheel model; Push Glide speed tracking is badly biased, so treat “command response” carefully. the 4 major comments →

arxiv 2607.10815 v1 pith:3MSREZFX submitted 2026-07-12 cs.RO

Learning Roller-Skating Motions of Humanoid Robots Based on Adversarial Motion Priors

Yunkang Cheng , Yutong Wu , Menghan Li , Shihe Zhou , Mingguo Zhao This is my paper
classification cs.RO
keywords humanoid robotroller-skatingreinforcement learningadversarial motion priorpassive wheelsmotion retargetingPump GlidePush Glide
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.

Humanoid roller skating is hard because the robot must keep whole-body balance while rolling on passive wheels that supply no drive torque of their own. This paper shows that two human skating styles—Pump Glide, a symmetric open-and-close skate pattern, and Push Glide, alternating push-off and glide—can each be learned as a separate policy from motion-capture demonstrations. Human data are retargeted to the robot, then used as style priors in adversarial motion-prior training, while task rewards handle velocity, posture, and contact safety, and skate wheels are modeled as sliced cylinders for stable rolling contact in simulation. Simulation and real-robot trials report sustained skating, velocity response, and gait-specific support timing. A sympathetic reader cares because this is a concrete route from human demonstration to passive-wheel humanoid locomotion without hand-coding a full phase machine for each gait.

Core claim

An AMP-based reinforcement learning pipeline—with independent reference datasets, policies, and reward architectures for Pump Glide and Push Glide, plus a nine-slice cylindrical collision model of passive skate wheels—can produce sustained humanoid roller-skating motions that appear in real-robot trials and show measurable velocity response and support-phase behavior.

What carries the argument

Adversarial motion prior (AMP) training with PPO: a discriminator scores short state transitions against retargeted demonstration clips so the policy inherits gait style, while gait-specific task rewards constrain velocity tracking, posture, and contact timing; passive wheels are simulated as nine narrow cylinders to keep rolling contact stable without false lateral support.

Load-bearing premise

The training stand-in for real skate wheels—nine thin cylinders plus randomized friction—has to be close enough to real rolling contact that the learned balance and style still work on the physical robot.

What would settle it

Deploy the same trained policies on the physical passive-skate humanoid under the reported velocity commands: if the robot cannot sustain Pump Glide open–close cycles or Push Glide alternating support, or if completion rate, torso tilt, and support timing diverge sharply from the simulation metrics without major retuning, the claimed learning–simulation–deployment pipeline fails.

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

If this is right

  • Distinct skating propulsion styles can be encoded by separate AMP pipelines without shared hand-crafted phase rules.
  • Pump Glide policies can maintain periodic foot opening–closing and long-horizon forward travel under velocity commands.
  • Push Glide policies can produce alternating support timing and command-sensitive forward speed on passive wheels.
  • A sliced-cylinder wheel model is usable enough for training that real-robot skating trials become feasible.
  • Passive-wheel humanoid locomotion can be cast as a demonstration-driven style-plus-task problem rather than pure model-based trajectory design.

Where Pith is reading between the lines

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

  • The same retarget–AMP–task-reward split could transfer to other underactuated human mobility devices with rolling contact, without redesigning phase machines from scratch.
  • The reported command-to-speed gain bias points to closed-loop speed correction or wheel-model identification as the next practical bottleneck for outdoor use.
  • If contact-model mismatch is the main sim-to-real limiter, refining wheel geometry or identifying friction online may improve tracking without discarding the learned gait priors.
  • Independent policies per gait imply a later need for a selector or multi-gait prior if continuous switching between Pump and Push is required.

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 / 7 minor

Summary. This paper proposes an AMP-PPO reinforcement learning pipeline for passive-wheel humanoid roller skating on a Booster T1 retrofitted with free-rolling skates. The authors introduce a 9-slice cylindrical collision model for skate wheels, retarget human mocap of two gaits (Pump Glide and Push Glide) via GMR into separate AMP reference datasets, and train independent policies with gait-specific task rewards and discriminators. Simulation studies report velocity-command sweeps, long-horizon tracking, and support-phase timing; real-robot trials provide qualitative snapshots of both gaits. The claimed contributions are (1) the sliced-wheel simulation method, (2) the demonstration-to-policy AMP pipeline for the two gaits, and (3) sim/real validation of sustained skating with velocity and support analysis.

Significance. Passive-wheel humanoid skating is a genuine underactuated contact-rich problem that sits between bipedal walking and wheeled mobility; demonstrating two stylistically distinct gaits with AMP priors and real-robot trials is a useful systems contribution for the wheeled-legged and humanoid RL communities. Strengths include a careful geometric analysis of wheel collision options (extra-volume fraction, roll-support angle, and training-throughput trade-off for the 9-slice model), explicit separation of style (AMP) from task objectives, and gait-specific contact-timing curricula for Push Glide. If the command-tracking and sim-to-real claims hold under tighter evaluation, the work would be a solid reference for passive-wheel humanoid locomotion. The result is incremental relative to prior AMP locomotion and recent skating/skateboarding RL papers, but the dual-gait passive-skate focus and wheel-modeling detail are concrete additions.

major comments (4)
  1. [§5.2 Table 2 / Fig. 9] §5.2, Table 2 and Fig. 9: Push Glide is presented as achieving “command-sensitive forward speed,” yet the stable-window actual speeds are systematically much higher than commanded (0.10→0.366 m/s, 0.50→1.594 m/s; errors 0.266–1.094 m/s). This is a large gain bias, not a small tracking residual. §6 attributes it to passive-wheel coupling and Isaac/MuJoCo mismatch, but Contribution 2–3 and §3.2 frame the task as operator velocity-command tracking (ct). Without closed-loop correction, gain calibration, or a clear statement that only monotonic sensitivity—not tracking accuracy—is claimed, the controllable-skating claim for Push Glide is overstated and reduces toward open-loop style generation.
  2. [Abstract vs §5] Abstract states that simulation experiments evaluate “gait quality, velocity tracking, turning, and gait-specific reward ablations.” §5 reports velocity sweeps, a long-horizon Pump Glide profile, and Push Glide support timing, but does not present turning results or reward-ablation studies. Either the missing experiments must be added with quantitative metrics, or the abstract and contribution framing must be narrowed to what is actually shown. As written, the paper promises evaluations that are load-bearing for the “independent reward architectures” claim and are not delivered.
  3. [§3.1, §5.2, §6] §3.1 and §6: The 9-slice cylinder model is central (Contribution 1) and is justified by geometry and throughput, but Push Glide evaluation is moved to MuJoCo while training is in Isaac Lab, and real-robot contact is a third regime. The manuscript admits mismatch as a source of velocity error yet still uses those runs as primary validation of command response and support timing. A load-bearing fix is needed: either quantify cross-simulator and sim-to-real contact fidelity (rolling radius, lateral friction, support height) for the adopted 9-slice model, or restrict claims to “sustained style under approximate rolling contact” and demote command-tracking claims accordingly.
  4. [§5.1–5.2 real-robot trials] §5 real-robot material (Figs. 5 and 8) is almost entirely qualitative snapshots, while completion rate, travel distance, torso tilt, and velocity error are reported only in simulation (and only thoroughly for Pump Glide). For a systems paper whose third contribution is “simulation and real-robot validation,” at least a minimal quantitative real-robot protocol (distance sustained, fall rate, approximate speed band, friction condition) is needed so that real-robot claims are not carried solely by images.
minor comments (7)
  1. [Table 1 / §4] Table 1 reward keys are listed without equations or precise definitions for several Push Glide terms (wheel_air_time_ratio, support_leg_switch_reward, wheels_spinning_reward, penalize_single_leg_ahead). A short appendix defining each ϕ_i would make the independent reward architectures reproducible.
  2. [§4 AMP features] §4: AMP state dimension is given as zm_t ∈ R^300 for a 5-frame stack, but the per-frame feature breakdown (which joints, which key points, units) is not specified. Please list the feature vector composition.
  3. [§3.1] §3.1: Wheel radius is written “R = 32 mm” correctly in places, but the sphere-width comparison text is clear; still, units should be checked consistently (one line earlier uses “32 mm” without issue—verify no “m” typos in camera-ready).
  4. [Fig. 3 / §3.1] Fig. 3 throughput plot is useful; please state whether “9-slice” is fixed for all later experiments and whether real-robot wheels match the 23.13 mm width assumption used in η_extra and ϕ_edge.
  5. [§2.2] Related work cites several concurrent arXiv skating/humanoid papers; ensure final versions and page numbers are updated, and clarify how SKATER [21] differs methodologically from the present Pump Glide policy (AMP vs hand-crafted/curriculum rewards).
  6. [§5.1] §5.1 completion rates ~0.77–0.81 over ~16 s trials: define failure (fall, timeout, foot collision) explicitly so the metric is interpretable.
  7. [Throughout] Minor prose/spacing issues appear throughout (e.g., missing spaces in “postureregulation,” “PumpGlideskating,” concatenated words in §2). A careful copy-edit pass is needed.

Circularity Check

0 steps flagged

No significant circularity: empirical AMP-PPO skating policies are trained from external mocap and hand-designed task rewards, not derived by tautology.

full rationale

This is an engineering/RL systems paper, not a first-principles derivation that claims to predict a quantity forced by its own fit. Human Pump/Push Glide mocap is collected and retargeted (GMR), then used as AMP reference distributions; task rewards (velocity tracking, foot spacing, support switching, air-time curricula, etc.) are explicit design choices with stated weights (Table 1) and mix coefficients (0.40/0.60 and 0.45/0.55). The sliced-cylinder wheel model is an approximation chosen for throughput/fidelity tradeoff, not a uniqueness theorem. Reported outcomes (completion rates, foot-separation rhythm, support timing, real-robot snapshots, and the Push Glide speed bias in Table 2/Fig. 9) are experimental measurements that can fail; the paper even admits velocity-tracking mismatch as a limitation. No step reduces a claimed prediction to a fitted input by construction, imports uniqueness from overlapping authors, or renames a known result as a forced derivation. Self-citations are standard AMP/DeepMimic/PPO/Isaac Lab background and are not load-bearing uniqueness claims. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 2 invented entities

The central claim rests on standard RL/control machinery plus many hand-chosen training knobs and modeling approximations. Load-bearing domain assumptions are passive-wheel dynamics, retargeting fidelity, and AMP as a style constraint. Free parameters dominate: reward weights, AMP coefficients, curricula, and the 9-slice wheel design. No new physical entity is postulated; the sliced-cylinder wheel is an engineering approximation without independent external validation beyond the authors' throughput/stability arguments.

free parameters (8)
  • Number of cylinder slices per wheel
    Chosen as 9 by throughput/fidelity tradeoff (Fig. 3); directly shapes contact geometry used for all training.
  • Task reward weights (Table 1, gait-specific)
    Dozens of hand-set weights (e.g., track_lin_vel_x_exp=12.0, feet_too_near, support_leg_switch_reward) define what 'success' means.
  • AMP vs task reward mix
    Pump: 0.40 AMP + 0.60 task; Push: 0.45 AMP + 0.55 task—chosen coefficients that control style vs task tradeoff.
  • AMP reward scale c_amp and tanh gain η
    c_amp=2.0, η=0.4, λ_gp=10 set discriminator reward magnitude and gradient penalty.
  • Push Glide contact-timing curriculum schedules
    Wheel air-time penalty 1.25→0.9 over 0–22000 steps; support-switch reward 0.2→1.0 by 32000 steps—hand-designed schedules.
  • Command curricula and velocity ranges
    Pump [0.2,0.7] m/s curriculum; Push [0,2] m/s and yaw ±0.2 rad/s resampling every 12 s—chosen task distributions.
  • Gait-specific action ranges / PD low-level controller gains
    Policy outputs scaled joint targets at 50 Hz PD; gains and ranges are free control parameters (partially randomized in training).
  • Domain randomization ranges
    Friction, mass, CoM, PD gains, wheel damping randomized; ranges not fully tabulated but affect reported robustness.
axioms (6)
  • domain assumption Passive skate wheels provide no drive torque; propulsion arises only from leg motion and wheel-ground friction (Iω̇ from tangential contact).
    Core task model in §1 and §3.1; if wheels were actuated or friction model wrong, the learned skill class changes.
  • ad hoc to paper A 9-slice narrow-cylinder collision model is an adequate proxy for real narrow-wheel rolling contact for training and evaluation.
    §3.1 geometric analysis rejects sphere/STL options; fidelity is argued but not independently validated against real contact forces.
  • domain assumption GMR retargeting from human mocap yields reference states that preserve skating style useful for AMP on the T1 morphology.
    §4 pipeline; morphology/joint-limit mismatch is acknowledged as motivation for AMP over strict tracking.
  • domain assumption Wasserstein AMP discriminators on 5-frame state features constrain gait style without requiring phase-aligned trajectory tracking.
    Standard AMP assumption from Peng et al., adopted in §4 with paper-specific features and losses.
  • standard math PPO with the stated observation/action spaces can optimize the combined AMP+task return under Isaac Lab passive-wheel simulation.
    Standard deep RL practice; no new optimization theory claimed.
  • domain assumption Real-robot trials on the modified T1 are comparable enough to simulation to support validation claims despite model mismatch.
    §5–6; authors note Isaac/MuJoCo and real-contact mismatch as sources of velocity error.
invented entities (2)
  • Sliced-cylinder passive skate-wheel collision model (9 cylinders) no independent evidence
    purpose: Provide stable rolling contact with reduced false lateral support versus full-sphere or mesh proxies during massively parallel training.
    Engineering approximation introduced for this platform; independent_evidence false because validation is internal (throughput, qualitative rolling stability), not external contact measurements.
  • Gait-specific dual AMP-PPO policies (Pump Glide / Push Glide) no independent evidence
    purpose: Separate style priors, rewards, and policies for two distinct propulsion mechanisms on the same robot.
    Learned controllers are the paper's product, not postulated physical entities; evidence is the reported sim/real behaviors.

pith-pipeline@v1.1.0-grok45 · 13896 in / 4150 out tokens · 60547 ms · 2026-07-14T09:02:38.598334+00:00 · methodology

0 comments
read the original abstract

Humanoid roller-skating is difficult because the robot must coordinate whole-body balance, rolling contacts, and velocity-dependent posture regulation. This paper presents an adversarial motion prior based reinforcement learning framework for two humanoid roller-skating gaits: Pump Glide skating and Push Glide skating. The two gait datasets are collected independently through motion capture and retargeted to the humanoid robot separately. The retargeted data are then smoothed and resampled into reference motion states for AMP training. The two gaits are learned by independent AMP training pipelines with separate reference datasets, separate policies, and independent reward architectures. Simulation experiments are designed to evaluate gait quality, velocity tracking, turning, and gait-specific reward ablations.

Figures

Figures reproduced from arXiv: 2607.10815 by Menghan Li, Mingguo Zhao, Shihe Zhou, Yunkang Cheng, Yutong Wu.

Figure 1
Figure 1. Figure 1: T1 humanoid skating platform, passive wheel modification, and system overview. into actuated-wheel and passive-wheel systems; this paper focuses on the latter, where propulsion is generated by leg motion and wheel-ground friction rather than wheel motors. The goal is to realize humanoid roller-skating motions on this passive-wheel system, not only in simulation but also in real-robot trials. This goal make… view at source ↗
Figure 2
Figure 2. Figure 2: Candidate collision models for skate wheels [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training throughput under different numbers of cylindrical slices the sphere collision width along the wheel axis is 2R = 64 mm, about 2.77 times the real wheel width. If a sphere-band is used as an upper-bound approximation of the true rounded wheel, the fraction of the full-sphere volume lying outside the true wheel width is ηextra = 1 − π(R2w − w 3/12) 4πR3/3 ≈ 48.2%. (1) Another error is the roll suppo… view at source ↗
Figure 4
Figure 4. Figure 4: AMP-based training framework for humanoid roller-skating gaits Both policies observe base angular velocity, projected gravity, velocity com￾mand, non-wheel joint states, and previous actions. The value network addition￾ally observes root linear velocity and foot contact indicators. The policy outputs target positions for 21 non-wheel joints only; these targets are scaled by gait￾specific action ranges and … view at source ↗
Figure 5
Figure 5. Figure 5: shows that the robot completes opening, maximum spacing, closing, and minimum spacing phases within one cycle in real-robot trials. During forward skating, the upper body remains stable and the leg motion is symmetric. The rhythm mainly comes from the AMP style reward, while stability is bounded by task rewards [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Pump Glide results under different forward velocity commands. Left: velocity, completion, and posture metrics; right: lateral foot separation Long-horizon velocity profile. The same policy tracks a 100 s profile that in￾creases from 0.1 m/s to 0.4 m/s and then decreases to zero [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Velocity response and lateral foot separation under a 100 s profile [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Real-robot Push Glide trials 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Time after command onset (s) 0.25 0.50 0.75 1.00 1.25 1.50 1.75 Filtered forward speed in yaw frame (m/s) 0.1 m/s actual 0.2 m/s actual 0.3 m/s actual 0.4 m/s actual 0.5 m/s actual 0.1 m/s 0.2 m/s 0.3 m/s 0.4 m/s 0.5 m/s 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 Stable speed (m/s) command stable actual [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Push Glide speed response under five commands [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Push Glide support phase and torso attitude may contribute to velocity-tracking errors in both Pump Glide and Push Glide. More accurate wheel-contact modeling, system identification, and closed-loop speed control remain future work. Acknowledgements This research was supported by STI 2030-Major Projects 2021ZD0201402, Beijing Natural Science Foundation (L243004), National Innovation and En￾trepreneurship … 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

23 extracted references · 2 canonical work pages

  1. [1]

    ACM Transactions on Graphics 37(4), 1–14 (2018)

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

    2018

  2. [2]

    ACM Transactions on Graphics 40(4), 1–15 (2021)

    Peng,X.B.,Ma,Z.,Abbeel,P.,Levine,S.,Kanazawa,A.:AMP:Adversarialmotion priors for stylized physics-based character control. ACM Transactions on Graphics 40(4), 1–15 (2021)

    2021

  3. [3]

    arXiv preprint arXiv:1707.06347 (2017)

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

    Pith/arXiv arXiv 2017

  4. [4]

    In: Proceedings of the 5th Confer- ence on Robot Learning, pp

    Rudin, N., Hoeller, D., Reist, P., Hutter, M.: Learning to walk in minutes using massively parallel deep reinforcement learning. In: Proceedings of the 5th Confer- ence on Robot Learning, pp. 91–100 (2022)

    2022

  5. [5]

    4565–4573 (2016) 12 Y

    Ho,J.,Ermon,S.:Generativeadversarialimitationlearning.In:AdvancesinNeural Information Processing Systems, pp. 4565–4573 (2016) 12 Y. Cheng et al

    2016

  6. [6]

    arXiv preprint arXiv:2510.02252 (2025)

    Araujo,J.P.,Ze,Y.,Xu,P.,Wu,J.,Liu,C.K.:Retargetingmatters:Generalmotion retargeting for humanoid motion tracking. arXiv preprint arXiv:2510.02252 (2025)

    arXiv 2025

  7. [7]

    In: IEEE International Conference on Robotics and Automation, pp

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

    2003

  8. [8]

    Autonomous Robots 40, 429–455 (2016)

    Kuindersma, S., Deits, R., Fallon, M., Valenzuela, A., Dai, H., Permenter, F., Koolen, T., Marion, P., Tedrake, R.: Optimization-based locomotion planning, es- timation, and control design for the Atlas humanoid robot. Autonomous Robots 40, 429–455 (2016)

    2016

  9. [9]

    In: Advances in Service and Industrial Robotics, pp

    Ficht, G., Behnke, S.: Direct centroidal control for balanced humanoid locomotion. In: Advances in Service and Industrial Robotics, pp. 242–255. Springer (2022)

    2022

  10. [10]

    In: Proceedings of CLAWAR 2024, pp

    Thibault, W., Melek, W., Mombaur, K.: Learning velocity-based humanoid lo- comotion: massively parallel learning with Brax and MJX. In: Proceedings of CLAWAR 2024, pp. 278–283. Springer (2024)

    2024

  11. [11]

    arXiv preprint arXiv:2404.05695 (2024)

    Gu, X., et al.: Humanoid-Gym: Reinforcement learning for humanoid robot with zero-shot sim-to-real transfer. arXiv preprint arXiv:2404.05695 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    arXiv preprint arXiv:2409.07846 (2024)

    Thibault, W., Rajendran, V., Melek, W., Mombaur, K.: Learning skateboarding for humanoid robots through massively parallel reinforcement learning. arXiv preprint arXiv:2409.07846 (2024)

    Pith/arXiv arXiv 2024

  13. [13]

    In: Proceedings of CLAWAR 2024, pp

    Rajendran, V., Mombaur, K.: Humanoids operating mobility devices designed for humans: experiments on a Segway. In: Proceedings of CLAWAR 2024, pp. 219–224. Springer (2024)

    2024

  14. [14]

    IEEE Robotics and Automation Letters 4(2), 2116–2123 (2019)

    Bjelonic, M., Bellicoso, C.D., de Viragh, Y., Sako, D., Tresoldi, F.D., Jenelten, F., Hutter, M.: Keep Rollin’–Whole-body motion control and planning for wheeled quadrupedal robots. IEEE Robotics and Automation Letters 4(2), 2116–2123 (2019)

    2019

  15. [15]

    Frontiers in Robotics and AI (2026)

    Petri, R., et al.: Passive wheeled-legged robots: A review. Frontiers in Robotics and AI (2026). doi:10.3389/frobt.2026.1857985

    work page doi:10.3389/frobt.2026.1857985 2026

  16. [16]

    In: IEEE/RSJ International Conference on Intelligent Robots and Systems, pp

    Todorov, E., Erez, T., Tassa, Y.: MuJoCo: A physics engine for model-based con- trol. In: IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033 (2012)

    2012

  17. [17]

    Biology Open 8(4), bio037713 (2019)

    Chen, J., Xu, K., Ma, H., Ding, X.L.: Motion characteristics of human roller skat- ing. Biology Open 8(4), bio037713 (2019). doi:10.1242/bio.037713

    work page doi:10.1242/bio.037713 2019

  18. [18]

    arXiv preprint arXiv:2402.16796 (2024)

    Cheng, X., et al.: Expressive whole-body control for humanoid robots. arXiv preprint arXiv:2402.16796 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    arXiv preprint arXiv:2406.10454 (2024)

    Fu, Z., et al.: HumanPlus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454 (2024)

    Pith/arXiv arXiv 2024

  20. [20]

    arXiv preprint arXiv:2504.14305 (2025)

    Shi, X., et al.: Adversarial locomotion and motion imitation for humanoid policy learning. arXiv preprint arXiv:2504.14305 (2025)

    arXiv 2025

  21. [21]

    arXiv preprint arXiv:2601.04948 (2026)

    Gu, J., Yuan, F., Shi, W., Huang, T., Zhang, H., Zhang, X., Wang, Y., Gao, W., Zhang, S.: SKATER: Synthesized kinematics for advanced traversing efficiency on a humanoid robot via roller skate swizzles. arXiv preprint arXiv:2601.04948 (2026)

    arXiv 2026

  22. [22]

    arXiv preprint arXiv:2603.18408 (2026)

    Wang, H., Fang, Z., Hanna, J., Xiong, X.: Efficient and versatile quadrupedal skating: Optimal co-design via reinforcement learning and Bayesian optimization. arXiv preprint arXiv:2603.18408 (2026)

    arXiv 2026

  23. [23]

    arXiv preprint arXiv:2511.04831 (2025)

    Mittal, M., et al.: Isaac Lab: A GPU-accelerated simulation framework for multi- modal robot learning. arXiv preprint arXiv:2511.04831 (2025)

    Pith/arXiv arXiv 2025