StableMimic: Smooth Human-Like Recovery for Humanoid Motion Tracking - Learning Beyond the Tracking Distribution for Structured Post-Fall Behavior

Chunying Li; Jinglei Nie; Ming Huang; Ruofei Liu; Shuxiang Guo; Weihao Wu

REVIEW 3 major objections 5 minor 23 references

StableMimic: Smooth Human-Like Recovery for Humanoid Motion Tracking - Learning Beyond the Tracking Distribution for Structured Post-Fall Behavior

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a single proprioception-gated policy can track humanoid motion, recover from falls in a human-like way, and resume the command without an external trigger or get-up reference.

desk verdict A genuinely useful unified tracking+recovery policy, with a real but addressable gap: the proprioceptive gate is only tested in-distribution and the stats are seedless. read the letter →

arxiv 2608.02385 v1 pith:UROPGVXP submitted 2026-08-03 cs.RO

Weihao Wu , Ming Huang , Ruofei Liu , Jinglei Nie , Shuxiang Guo , Chunying Li This is my paper · ORCID

classification cs.RO

keywords humanoidmotiontrackingpost-fallrecoverymixture-of-expertsproprioceptivegatingreinforcementlearningsuccessor-staterewardget-upreferenceUnitreeG1

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

StableMimic tries to establish that a humanoid motion tracker can be trained to handle states outside its normal tracking distribution — specifically falls — within the same policy, so the robot recovers in a human-like, bounded way and resumes its command. It claims that perturbed training resets around human get-up demonstrations teach structured recovery, while a proprioception-only gate continuously blends a tracking expert and a recovery expert, avoiding a hard switch or runtime reference. If correct, one deployable policy can replace separate tracking and stand-up systems, and robots pushed over during dance or teleoperation will not thrash or chase unreachable commands. The paper reports the lowest errors on all four tracking metrics over complete LAFAN1 dance sequences among five methods, 100/100 recovery in matched push-to-fall trials with the lowest values on six of seven post-fall motion and load measures, and qualitative real-robot recovery on a Unitree G1.

What carries the argument

The central object is a proprioception-gated dual-expert mixture-of-experts policy. Two experts map the same command-conditioned observation history to joint-target action means — one for tracking the live command, one for recovery shaped by hidden get-up references — while a soft gate takes only the proprioceptive history (base angular velocity, projected gravity, joint states, previous action) and computes a softmax blend of the two experts' outputs, so there is no discrete action switch. Training uses perturbed get-up resets around multiple human get-up trajectories to sample a recovery support beyond the command-tracking tube, and a hidden successor-state reward rewards approaching the n

What would settle it

Run the deployed policy on falls whose initial states lie outside the sampled support used in training — for example, push directions not in {+x, -x, +y, -y}, or a fall from a small ledge — and record whether the recovery expert's gate weight stays high during ground contact and whether tracking resumes in the 1.5-s transition; a single such fall with sustained tracking-expert weight during low-height contact states would refute the claim that proprioception alone distinguishes recovery states from tracking states.

Watch

Extended reading notes

Core claim

StableMimic's central claim is that extending a tracking policy's training support into perturbed neighborhoods of human get-up references, and routing post-fall states to a dedicated recovery expert through a proprioceptive gate, yields a single policy that both tracks a continuously advancing whole-body command and produces human-reference-shaped recovery without any runtime get-up reference, recovery command, or policy switch. In simulation, the gated dual-expert policy attains the best results on all four tracking metrics over the complete LAFAN1 dance subset, recovers in 100/100 matched push-to-fall trials, and achieves the lowest values on six of seven post-fall motion and load measure

Load-bearing premise

The load-bearing premise is that a four-frame history of the robot's own body readings alone is enough, at deployment, to tell when the robot is in a post-fall recovery state rather than still tracking a commanded pose, because the gate's routing labels are available only during training and the command/reference is never given to it at runtime; the paper itself limits this claim to the sampled perturbation neighborhoods around get-up references.

Editorial extensions

If this is right

If correct, one deployable policy can track motion, fall, recover, and resume the command without external fall detection, a get-up reference, a recovery command, or a policy switch.
Training beyond the tracking distribution does not degrade tracking: the MoE policy achieves lower MPBPE and MJAE than a larger Single-MLP ablation and a tracking-only baseline on the complete LAFAN1 dance subset.
Post-fall motion and actuator load are reduced under the matched protocol: 100/100 recovery versus 98/100 for the Single-MLP and 0/100 for the tracking-only baseline, with lower limb speed, limb travel, joint speed, actuator torque, and positive energy.
The same exported ONNX policy runs on real Unitree G1 hardware at 50 Hz, and both dance and constant-standing-reference deployments show bounded limb motion, autonomous recovery, and task resumption.
Because the method uses a common observation/action interface, it can augment other RL trackers with compatible spaces, extending structured post-fall behavior to teleoperation-oriented systems.

Reading between the lines

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

Editorial extension: because the gate sees no command or reference, its ability to distinguish a fall from an intentionally low commanded pose is only as good as the four-frame proprioceptive history; a natural test is commanding a crouch profile and checking that tracking expert weight stays dominant.
Editorial extension: the recovery guarantee is explicitly bounded by the sampled perturbation neighborhoods around get-up references, so falls outside that support (new push directions, uneven terrain, ledges) may not recover in the demonstrated shape — a closed-loop curriculum that grows the support could be the next step.
Editorial extension: since reference identity and phase are hidden from the deployed Actor, the policy may be learning a general 'get upright' strategy rather than memorising specific get-up trajectories; evaluating on held-out get-up styles not present in the library would test this directly.
Editorial extension: the approach could be combined with fall-prediction or impact-mitigation policies: a predicted fall could trigger a protective pre-impact controller that hands over smoothly to the recovery expert via the same soft gate mechanism.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

StableMimic is a solid, useful paper, and the main experimental claims are supported as far as the paper goes. The genuinely new thing is the single deployable policy that does both tracking and human-shaped post-fall recovery, with a proprioceptive gate that blends two experts and no runtime get-up reference, recovery command, or hard switch. That packaging removes a real failure mode of tracking-only policies, and the matched LAFAN1 tracking plus 100-trial recovery protocol is the right way to demonstrate it. I read the training curves and the recovery numbers as consistent with the intended expert separation; the Single-MLP comparison is fairer than most because it has a larger actor yet does worse. The authors are also honest about scope — they say states outside R_delta are outside the claim, and they call the hardware results qualitative. Credit where due.

The soft spots are real but not fatal. First, no code or data, and single deterministic runs. The tables have no seed variance or error bars, so 100/100 could be a point estimate from one lucky run; for an RL paper this is an addressable gap, not a flaw in the method. Second, the gate-transfer premise. The gate is trained with rollout-regime labels, and the push protocol (525–575 N for 0.2 s) likely visits states beyond the reset perturbations (linear velocity ±0.5 m/s, angular ±0.52 rad/s). The paper explicitly excludes such states. So the claim that the proprioceptive gate routes correctly outside R_delta is not demonstrated; 100/100 under one scripted protocol is thin evidence for a distributional claim. I would not call this circular — using get-up references in the reset/reward is standard reward shaping — but it is missing evidence. Third, the baselines are implemented by reference, which makes fairness hard to verify without code; the matched protocol helps, but only partly.

Who is this for? Anyone working on humanoid tracking, teleoperation, or fall recovery in sim-to-real RL. It deserves a serious referee. My recommendation: send to peer review, and ask the authors to release code/data and rerun with seeds and error bars, plus a couple of disturbance protocols outside R_delta and gate-routing statistics. If those land, the paper is close to acceptance.

Referee Report

3 major / 5 minor

Summary. The paper presents StableMimic, a single proprioception-gated dual-expert policy for humanoid command tracking and post-fall recovery. Training augments the nominal tracking task with perturbed resets around human get-up references and a hidden successor-state reward; the deployed actor receives only the live command and proprioceptive history, with no get-up reference, recovery command, or external policy switch. Experiments on the Unitree G1 in MuJoCo report lower tracking errors on the LAFAN1 dance subset than a Single-MLP ablation and two prior systems, 100/100 recovery on a matched 100-push protocol, and qualitative real-hardware confirmation.

Significance. If the results hold, the paper makes a practically valuable contribution: it demonstrates a plausible path toward a single deployable policy that resumes uninterrupted command tracking after a fall, without external triggering or reference lookup, while improving post-fall motion and load metrics. The paper's strengths include a matched protocol across methods, an explicit training/deployment information-boundary table, a larger-capacity Single-MLP ablation, honest scoping statements, and careful differentiation between measured results and qualitative hardware demos. The central weakness is that the main claim about behavior during unplanned falls is evaluated under pushes that likely produce states outside the paper's declared recovery support, so the gate-transfer premise is not yet established.

major comments (3)

[§III-B / Table V / §IV-A] The recovery support Rδ is explicitly an empirical set of perturbed neighborhoods around get-up references, with reset velocity bounds |v_lin|≤0.5 m/s, |ω|≤0.52 rad/s, and joint noise ±0.1 rad; the paper states 'states beyond this sampled support are outside our claim.' The push protocol applies 525–575 N for 0.2 s to a G1; a simple momentum estimate (Δv ≈ FΔt/m) already exceeds 0.5 m/s for plausible G1 masses, and the resulting contact configurations are unlikely to lie inside Rδ. Thus the main recovery evaluation exercises states the paper self-excludes. Because the gate in §III-C(a) was trained only on privileged rollout-regime labels within Rδ, its routing accuracy on push-induced states is unknown. The 100/100 result is from one deterministic protocol without error bars, so it cannot by itself establish robustness. Please report the empirical distribution of push-induced states rela
[§IV-A(c)] Recovery success is defined only as 'a structured get-up and tracking resumption without resetting the policy or command.' The fall onset is quantified (pelvis height <0.50 m or tilt >60°), but 'structured get-up' is not. This makes the 100/100 success number non-reproducible and weakens the comparison across methods. Please specify an operational success criterion (e.g., pelvis height above a threshold, task tracking error below a bound for a duration, and no fall recurrence within the horizon).
[Tables III and IV] All comparisons are based on a single training run per method. RL policies have run-to-run variance; differences such as StableMimic vs Single-MLP MPBPE (28.53 vs 32.66 mm) or recovery success (100/100 vs 98/100) may not be significant. Without multiple seeds (or otherwise showing variance bounds), the claim that StableMimic achieves the lowest tracking and recovery metrics is unsupported in a statistical sense. Please report means and standard deviations (or distributions) over at least 3–5 seeds.

minor comments (5)

[Fig. 4 and Fig. 2 inset] Figure 4 is labeled 'conceptual' and the inset of Fig. 2 is qualitative. Consider adding a measured example of gate weights or routing probabilities over a fall-recovery rollout to support the 'soft blending' narrative.
[§IV-A(a)] The number of LAFAN1 dance sequences used for evaluation and the total duration of the 'complete dance subset' are not stated. Please specify these to improve reproducibility.
[Eq. (3)] The temperature τ in the softmax gate is not reported. Please state its value or scheduling.
[§III-C(a)] The 'relative-height safeguard' is described only verbally. Please give the exact condition (e.g., h_reset_base ≥ 0.8 h_cmd) and clarify how it interacts with gate targets for non-reset transitions.
[Appendix Table V] The reset ranges are reported per dimension; please clarify whether these are sampled uniformly and whether they apply to the root only or to all links/joints.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the results are measured against external baselines and the training objectives are not fitted to the test predictions.

full rationale

The paper's central claim is not circular. StableMimic trains a dual-expert policy with standard PPO (Eqs. 3 and 4); the get-up reference appears only in training resets, rewards, and Critic inputs, and is explicitly masked from the deployed Actor (`o^π_t = H^4(ocmd_t, 0_get, oprop_t, a_{t-1})`, Eq. 1). No parameter or metric is fitted to the reported test outcomes: the tracking numbers in Table III and the 100/100 recovery result in Table IV are measured on a fixed 100-disturbance protocol (§IV-A) that is not used to tune the policy. The gate's privileged rollout-regime labels (§III-C(a)) are training supervision, not a definition of the evaluation success; the push protocol (525–575 N) lies outside the training reset perturbations (Table V), so the recovery result is an independent empirical outcome rather than a construction. The paper's scope statement at §III-B, 'states beyond this sampled support are outside our claim', is an honest limitation about generalization, not a circular derivation. There is no load-bearing self-citation, no imported uniqueness theorem, and no renamed known result. The comparison against BeyondMimic, KungFuAthlete, BFM-Zero, and the Single-MLP ablation provides independent external checks.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

All entries are training/architecture design choices rather than physical laws. The central claim depends on the empirical support Rδ (Eq. 2) covering realized post-fall states and on the proprioceptive gate generalizing from privileged regime labels to deployment; both are treated as assumptions rather than proven. No free parameters are fitted to the evaluation outcomes; they are hand-selected hyperparameters, which is standard for RL but still means the results are only as good as the tuning.

free parameters (5)

Recovery reward coefficient multiplier = 2.5× λ_cmd
Eq. (4): hand-chosen to strengthen the recovery-phase reference term; affects the relative weight of get-up imitation vs regularization.
Gate auxiliary loss weights = CE 0.1; transition 4.0; consistency 0.01; alignment 0.01
Appendix Table V: hand-tuned weights for gate-target cross-entropy, transition alignment, and routing consistency; directly shape expert specialization and gate dynamics.
Reset perturbation ranges = x,y ±0.05 m; z ±0.01 m; roll/pitch ±0.1 rad; yaw ±0.2 rad; linear vel ±0.5/0.2 m/s; angular vel ±0.52/0.78 rad/s; joint
Table V: define the recovery support Rδ; if too narrow, real falls will land outside the trained distribution and the central claim fails.
Transition duration and error tolerance = 1.5 s transition; 2.0 s target-error tolerance
§III-B: thresholds governing when recovery ends and tracking resumes, and when failed recovery terminates the episode; influence success counts.
Failure-adaptive sampler hard/uniform mix = not specified (stated as 'equally')
§III-B: the mix between hard-frame and uniform sampling affects the density of the recovery support; exact value unstated.

assumptions (5)

standard math Clipped PPO with GAE converges to a performant policy for the shaped reward.
§III-D: standard RL convergence assumption; no formal guarantee.
domain assumption MuJoCo simulation with dynamics randomization faithfully represents Unitree G1 enough for sim-to-real transfer.
§IV-A/D: hardware results are qualitative; no quantitative transfer analysis.
domain assumption Retargeted LAFAN1 get-up references provide a valid human-shaped recovery target for G1.
§III-B/IV-A: the recovery reward and resets are defined by these references; no independent evidence they are physically optimal or fully cover post-fall states.
domain assumption Four-frame proprioceptive history is sufficient to infer recovery vs tracking regime at deployment.
§III-A/C: Actor sees oπt = H4(ocmd, 0get, oprop, at-1) with get-up features masked; if two regimes share proprioceptive signatures, gate generalization fails.
domain assumption The success definition ('structured get-up and tracking resumption') is a valid proxy for interaction safety.
§IV-A(c): motion/load metrics are used as safety evidence; the paper explicitly does not certify safety.

how reviews work

0 comments

Cite this review

Pith. "Pith review of StableMimic: Smooth Human-Like Recovery for Humanoid Motion Tracking - Learning Beyond the Tracking Distribution for Structured Post-Fall Behavior." pith.science (2026). https://pith.science/paper/UROPGVXP

@misc{pith2026260802385,
  author       = {Pith},
  title        = {Pith review of: StableMimic: Smooth Human-Like Recovery for Humanoid Motion Tracking - Learning Beyond the Tracking Distribution for Structured Post-Fall Behavior},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UROPGVXP}},
  note         = {Machine review of arXiv:2608.02385}
}

read the original abstract

Humanoid motion trackers perform reliably within learned tracking distributions, but falls can move the robot into low-height, contact-rich states from which an advancing command is temporarily unreachable. Tracking-only policies may chase infeasible references, producing rapid, large-amplitude limb corrections that increase risk to the robot and its surroundings. We present StableMimic, a unified tracker trained beyond the nominal tracking distribution. Perturbed resets around multiple human get-up references expose prone, supine, off-balance, and intermediate ground-contact states, shaping structured recovery that returns the robot to the trackable region. Because tracking and recovery occupy markedly different state--action distributions, StableMimic uses dedicated experts for each regime and a proprioceptive gate that continuously blends their actions. A hidden successor-state objective teaches human-reference-shaped recovery without exposing reference identity or phase to the deployed Actor; deployment requires no get-up reference, recovery command, trajectory retrieval, or external policy switch. On the complete retargeted LAFAN1 dance subset, StableMimic achieves the lowest errors on all four tracking metrics among five methods. Across 100 matched push-to-fall trials per method, it recovers in 100/100 and attains the lowest values on six of seven post-fall motion and load measures, supporting improved interaction safety under this protocol. Real Unitree G1 dance and standing-reference deployments qualitatively demonstrate bounded limb motion, autonomous recovery, and command resumption.

Figures

Figures reproduced from arXiv: 2608.02385 by the authors.

**Figure 1.** StableMimic overview. One policy tracks a continuously advancing whole-body command, routes fall-induced proprioceptive states toward recovery, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** StableMimic method overview. Both experts receive the live command and deployable observation history, while the gate uses proprioceptive history [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Recovery-reference learning and gate supervision. (A) Perturbed get-up states and a permanently masked reference define the successor-state reward [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Conceptual state–time view of the learned support. The blue curve and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Representative tracking–fall–recovery rollouts in MuJoCo. Top: [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Training dynamics over 30,000 iterations. StableMimic stabilizes episode length and reward earlier and reaches a lower policy standard deviation than Single-MLP. a) Overall post-fall performance: As reported in Table IV and [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Tracking and post-fall evaluation on the complete dance subset of LAFAN1. (a)–(c) Temporal tracking errors; (d) per-sample body-position versus [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Real-G1 deployment of dance (top) and standing-reference (bottom) policies, showing recovery and task resumption with bounded limb motion. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 3 linked inside Pith

[1]

FRASA: An End-to-End Reinforcement Learning Agent for Fall Recovery and Stand Up of Humanoid Robots,

C. Gaspard, M. Duclusaud, G. Passault, M. Daniel, and O. Ly, “FRASA: An End-to-End Reinforcement Learning Agent for Fall Recovery and Stand Up of Humanoid Robots,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), pp. 15994–16000, 2025

2025
[2]

Learning Humanoid Standing-up Control across Diverse Postures,

T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, and J. Pang, “Learning Humanoid Standing-up Control across Diverse Postures,” in Proc. Robotics: Science and Systems (RSS), 2025

2025
[3]

Learning Getting-Up Policies for Real-World Humanoid Robots,

X. He, R. Dong, Z. Chen, and S. Gupta, “Learning Getting-Up Policies for Real-World Humanoid Robots,” inProc. Robotics: Science and Systems (RSS), 2025

2025
[4]

BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,

Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu, “BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,”arXiv:2508.08241, 2025

arXiv 2025
[5]

SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control,

Z. Luo et al., “SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control,”arXiv:2511.07820, 2025

arXiv 2025
[6]

A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking,

Z. Lei, L. Cao, X. Wang, T. Qian, J. Liu, and X. Li, “A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking,”arXiv:2602.13656, 2026

arXiv 2026
[7]

BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning,

Y . Li et al., “BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning,” inProc. Int. Conf. Learning Representations (ICLR), 2026

2026
[8]

DeepMimic: Example- Guided Deep Reinforcement Learning of Physics-Based Character Skills,

X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “DeepMimic: Example- Guided Deep Reinforcement Learning of Physics-Based Character Skills,” ACM Transactions on Graphics, vol. 37, no. 4, Art. 143, 2018

2018

Show all 23 references

[9]

AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control,

X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control,”ACM Transac- tions on Graphics, vol. 40, no. 4, Art. 144, 2021

2021
[10]

Embrace Collisions: Humanoid Shadowing for De- ployable Contact-Agnostics Motions,

Z. Zhuang and H. Zhao, “Embrace Collisions: Humanoid Shadowing for De- ployable Contact-Agnostics Motions,” inProc. Conf. Robot Learning (CoRL), 2025

2025
[11]

SafeFall: Learning Protective Control for Humanoid Robots,

Z. Meng et al., “SafeFall: Learning Protective Control for Humanoid Robots,” arXiv:2511.18509, 2025

2025
[12]

Robot Crash Course: Learning Soft and Stylized Falling,

P. Strauch et al., “Robot Crash Course: Learning Soft and Stylized Falling,” in Proc. IEEE Int. Conf. Robotics and Automation (ICRA), 2026

2026
[13]

Unified Humanoid Fall-Safety Policy from a Few Demonstrations,

Z. Xu, Y . Li, K.-Y . Lin, and S. X. Yu, “Unified Humanoid Fall-Safety Policy from a Few Demonstrations,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), 2026

2026
[14]

Outrageously Large Neural Networks: The Sparsely-Gated Mixture- of-Experts Layer,

N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously Large Neural Networks: The Sparsely-Gated Mixture- of-Experts Layer,” inProc. Int. Conf. Learning Representations (ICLR), 2017

2017
[15]

Multi-Task Reinforcement Learning with Soft Modularization,

R. Yang, H. Xu, Y . Wu, and X. Wang, “Multi-Task Reinforcement Learning with Soft Modularization,” inAdvances in Neural Information Processing Systems, vol. 33, pp. 4767–4777, 2020

2020
[16]

Terrain-Adaptive Locomotion Skills Using Deep Reinforcement Learning,

X. B. Peng, G. Berseth, and M. van de Panne, “Terrain-Adaptive Locomotion Skills Using Deep Reinforcement Learning,”ACM Transactions on Graphics, vol. 35, no. 4, Art. 81, 2016

2016
[17]

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,”Science Robotics, vol. 5, no. 49, Art. eabb2174, 2020

2020
[18]

Acquiring Diverse Skills Using Curriculum Reinforcement Learning with Mixture of Experts,

O. Celik, A. Taranovic, and G. Neumann, “Acquiring Diverse Skills Using Curriculum Reinforcement Learning with Mixture of Experts,” inProc. 41st Int. Conf. Machine Learning, PMLR, vol. 235, pp. 5907–5933, 2024

2024
[19]

CMoE: Contrastive Mixture of Experts for Motion Control and Terrain Adaptation of Humanoid Robots,

S. Ma, H. Chen, Z. Xu, Y . Zhao, K. Wu, R. Yang, L. Zou, Z. Gan, and W. Ding, “CMoE: Contrastive Mixture of Experts for Motion Control and Terrain Adaptation of Humanoid Robots,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), 2026

2026
[20]

Between MDPs and Semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning,

R. S. Sutton, D. Precup, and S. Singh, “Between MDPs and Semi-MDPs: A Framework for Temporal Abstraction in Reinforcement Learning,”Artificial Intelligence, vol. 112, no. 1–2, pp. 181–211, 1999

1999
[21]

Asym- metric Actor Critic for Image-Based Robot Learning,

L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asym- metric Actor Critic for Image-Based Robot Learning,” inProc. Robotics: Science and Systems (RSS), 2018

2018
[22]

Proximal Policy Optimization Algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,”arXiv:1707.06347, 2017

2017 arXiv
[23]

Robust Motion In- Betweening,

F. G. Harvey, M. Yurick, D. Nowrouzezahrai, and C. Pal, “Robust Motion In- Betweening,”ACM Transactions on Graphics, vol. 39, no. 4, Art. 60, 2020

2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

FRASA: An End-to-End Reinforcement Learning Agent for Fall Recovery and Stand Up of Humanoid Robots,

C. Gaspard, M. Duclusaud, G. Passault, M. Daniel, and O. Ly, “FRASA: An End-to-End Reinforcement Learning Agent for Fall Recovery and Stand Up of Humanoid Robots,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), pp. 15994–16000, 2025

2025

[2] [2]

Learning Humanoid Standing-up Control across Diverse Postures,

T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, and J. Pang, “Learning Humanoid Standing-up Control across Diverse Postures,” in Proc. Robotics: Science and Systems (RSS), 2025

2025

[3] [3]

Learning Getting-Up Policies for Real-World Humanoid Robots,

X. He, R. Dong, Z. Chen, and S. Gupta, “Learning Getting-Up Policies for Real-World Humanoid Robots,” inProc. Robotics: Science and Systems (RSS), 2025

2025

[4] [4]

BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,

Q. Liao, T. E. Truong, X. Huang, Y . Gao, G. Tevet, K. Sreenath, and C. K. Liu, “BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion,”arXiv:2508.08241, 2025

arXiv 2025

[5] [5]

SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control,

Z. Luo et al., “SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control,”arXiv:2511.07820, 2025

arXiv 2025

[6] [6]

A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking,

Z. Lei, L. Cao, X. Wang, T. Qian, J. Liu, and X. Li, “A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking,”arXiv:2602.13656, 2026

arXiv 2026

[7] [7]

BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning,

Y . Li et al., “BFM-Zero: A Promptable Behavioral Foundation Model for Humanoid Control Using Unsupervised Reinforcement Learning,” inProc. Int. Conf. Learning Representations (ICLR), 2026

2026

[8] [8]

DeepMimic: Example- Guided Deep Reinforcement Learning of Physics-Based Character Skills,

X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “DeepMimic: Example- Guided Deep Reinforcement Learning of Physics-Based Character Skills,” ACM Transactions on Graphics, vol. 37, no. 4, Art. 143, 2018

2018