RL from Physical Feedback: Aligning Large Motion Models with Humanoid Control

Jiangxing Wang; Junpeng Yue; Sipeng Zheng; Weishuai Zeng; Xinrun Xu; Yuxuan Wang; Yu Zhang; Zepeng Wang; Ziluo Ding; Zongqing Lu

REVIEW 5 major objections 7 minor 4 cited by

RL from Physical Feedback: Aligning Large Motion Models with Humanoid Control

T0 review · 5 major / 7 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 RL fine-tuning of a text-to-motion generator with a binary physical-tracking reward lifts simulated humanoid tracking success from about 0.43 to 0.95 while preserving semantic alignment.

desk verdict A genuinely new RL-based T2M fine-tuning framework, but the headline success-rate gains are circular because the evaluation metric is exactly the training reward, and the paper's own ablation shows that metric is gameable. read the letter →

arxiv 2506.12769 v1 pith:4K43SCYC submitted 2025-06-15 cs.RO cs.LG

Junpeng Yue , Zepeng Wang , Yuxuan Wang , Weishuai Zeng , Jiangxing Wang , Xinrun Xu , Yu Zhang , Sipeng Zheng

show 2 more authors

Ziluo Ding Zongqing Lu

This is my paper

classification cs.ROcs.LG

keywords text-to-motiongenerationhumanoidcontrolreinforcementlearningfromphysicalfeedbackmotiontrackingpolicyGRPOsim-to-realtransfersemanticalignmentfeasibility

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 claims that text-to-motion generation models can be fine-tuned to produce motions that a humanoid robot can actually execute, without losing the semantic match to the text. To do this, the authors treat a pre-trained motion-tracking policy as a physical oracle: it simulates the generated motion on a Unitree G1 model and returns a binary success signal for whether the robot tracked it within 0.5 m of the reference. That signal becomes a reward for Group Relative Policy Optimization (GRPO), and a contrastive alignment module keeps the generated motion close to both the text description and ground-truth motion. On CMU and AMASS benchmarks, the fine-tuned generator lifts simulated tracking success from roughly 0.43/0.48 to 0.95/0.97 and 0.92/0.90 respectively, while multimodal distance stays near ground truth. The paper reports that the resulting motions deploy on a real Unitree G1, which is the concrete payoff of the approach.

What carries the argument

The load-bearing object is the frozen student motion-tracking policy trained in an Exbody2-style two-stage teacher–student setup on the AMASS dataset. It converts any retargeted motion into a scalar physical-feasibility signal—binary success if the robot tracks the reference with mean deviation below 0.5 m and no loss of balance—which is then used as the reward $R_{\text{tracking}}$ for GRPO fine-tuning of the LLM-based motion generator. The second mechanism, the alignment verification module, uses contrastive text and motion encoders (Eq. 6) to produce two regularizers: text alignment $R_{\text{TA}}$ and motion alignment $R_{\text{MA}}$ (Eq. 7), which keep the RL-optimized motions semantically faithful and kinematically close to ground truth. Together, the tracking reward biases the generator toward executable motions while the alignment terms stop the optimizer from collapsing to degenerate but trackable outputs.

What would settle it

Deploy the motions produced by RLPF-w/o align (tracking reward only, no alignment verification) on the Unitree G1 in the real world; if the standing-degenerate motions that score 0.99 Succ in IsaacGym fail to transfer or are trivially executable but carry no semantic content, then the binary tracking reward is gameable and the physical-feasibility signal is not a trustworthy oracle.

Watch

Extended reading notes

Core claim

In the paper's own terms, RLPF establishes that physical feasibility can be learned as a post-training objective for a large motion model rather than engineered into the generator. The discovery is that a frozen, generalizable motion-tracking policy, trained on AMASS and deployed in IsaacGym, provides a reward signal strong enough to reshape the generator's output distribution: optimizing the binary tracking success reward $R_{\text{tracking}} = \mathbb{I}(\mathrm{Succ}(\pi, m_i))$ with GRPO is sufficient to make generated motions trackable, and the addition of text and motion alignment verification (Eq. 7) prevents the semantic collapse that pure physical-reward optimization would otherwise cause. The reported numbers—a jump in Success Rate from 0.43/0.48 on the base model to 0.95/0.97 on CMU and 0.92/0.90 on AMASS in IsaacGym, with a smaller but consistent gain in zero-shot MuJoCo—are the paper's evidence that the joint optimization preserves semantic alignment while achieving physical executability.

Load-bearing premise

The load-bearing premise is that the pre-trained tracking policy, tested only in simulation, is a faithful proxy for what a real Unitree G1 can execute; if a generated motion can fool that policy into reporting success without being physically executable, the reported feasibility gains do not carry over to reality.

Editorial extensions

If this is right

Motions generated by RLPF can be dropped directly into a humanoid control stack, reducing the need for per-motion parameter tuning.
The same GRPO-plus-tracker recipe can be applied to any pretrained text-to-motion model, since the framework is agnostic to the specific generator architecture.
Zero-shot transfer from IsaacGym to MuJoCo suggests that motions optimized under one simulator retain trackability in another, which is a precondition for real-world sim-to-real transfer.
The alignment verification module keeps text-motion retrieval metrics close to ground truth, so the physical-feasibility gains do not come at the cost of instruction following.
Because the tracking reward is a learned policy rather than a hand-designed penalty, the approach scales to new motions without re-engineering physical constraints.

Reading between the lines

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

I extend the paper's reasoning: the binary success threshold at 0.5 m may be too coarse, since a motion could pass it while still containing foot sliding or penetration that the policy tolerates; a graded reward or multi-metric success criterion would give a stronger feasibility signal.
I infer from the paper's own limitation section that freezing the tracking policy caps what RLPF can teach; a natural extension is to jointly train the generator with an adaptable tracking policy so out-of-distribution motions remain learnable.
I read the RLPF-w/o align ablation as evidence that the tracking reward and alignment verification are not separable in practice: without the alignment terms, the optimizer exploits the reward by generating degenerate standing motions that score 0.99 success, so the reported feasibility gains should be interpreted as a property of the joint objective, not the tracking reward alone.
I suggest the same reward design transfers to other embodied generation tasks, such as text-to-grasp or text-to-navigation, wherever a frozen simulator policy can act as an oracle for executability.

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.

Desk Editor's Note

read the letter

The core idea here is new and worth taking seriously: fine-tune a large text-to-motion generator with GRPO, using a frozen humanoid tracking policy's binary success as the physical-feasibility reward, and add a contrastive alignment term to stop semantic collapse. That integration is not in the cited literature, and the paper is refreshingly honest about the failure mode it is trying to avoid—the RLPF-w/o align ablation in Tables 5 and 6 shows the tracking reward alone drives the model to emit standing sequences with 0.99 Succ while retrieval metrics collapse. That is real evidence that the alignment module is doing necessary work. The cross-simulator zero-shot transfer to MuJoCo and the Unitree G1 videos are also encouraging, and the ablations swapping in PHC are a sensible attempt to check policy dependence.

The problem is the central quantitative claim. Succ is both the training reward (Eq. 5) and the primary evaluation metric (Section 4.1), so the reported jump from ~0.45 to ~0.95 is, by construction, the model getting better at the exact thing it was trained to maximize. The paper itself demonstrates that a high Succ can come from physically trivial behavior—standing still is perfectly trackable. Adding the alignment term prevents that collapse, but it does not establish that Succ is a valid proxy for real-world physical feasibility; it only correlates with what the Exbody2-style student policy can track. The MuJoCo results are less circular, but they use the same frozen policy, so the bias carries over. The real-robot evidence is qualitative, no numeric tracking error or fall rates.

Other soft spots: no error bars anywhere; the cited physics-aware T2M baselines (PhysDiff, Reindiffuse) are discussed but never compared; FID degrades substantially after fine-tuning (e.g., 2.35 to 3.61 on CMU) without discussion; and no code is released. There is also a minor inconsistency—the paper describes both text-alignment and motion-alignment verification, but reports RLPF-MA and RLPF-TA as separate variants rather than combining them.

Who is this for? Researchers working at the intersection of text-to-motion and humanoid control, who will find the framework a useful starting point. The paper deserves a serious referee, but the referee should demand an independent physical-validity metric—a different tracking policy not used as reward, or a dynamics-based plausibility score—plus error bars and honest comparisons to the physics-aware methods already cited. My own verdict is skeptical, not because the idea is bad, but because the evidence, as presented, does not support the load-bearing claim.

Referee Report

5 major / 7 minor

Summary. The paper proposes RLPF, a framework for fine-tuning a pre-trained LLM-based text-to-motion generator with GRPO, using rewards from a frozen humanoid tracking policy (binary tracking success, Eq. 5) and an alignment verification module (Eq. 7) that preserves text/motion semantics. Experiments on CMU and AMASS report large gains in IsaacGym and MuJoCo Success Rate (up to about 0.95–0.99) over Base, SFT, and SFT-Filter, with MMDist near ground truth, plus a video of deployment on a Unitree G1. The central claim is that RLPF produces motions that are semantically aligned and physically executable by a real humanoid.

Significance. The idea of using a physics simulator's tracking policy as a reward source for text-to-motion generation is timely and could be practically valuable. The paper also correctly identifies and demonstrates, via the RLPF-w/o align ablation, that a pure tracking reward is gameable, and the alignment verification module is a sensible countermeasure. However, the current quantitative evidence does not establish the central claim: the primary evaluation metric, Success Rate, is the same binary signal used as the training reward, and the paper's own ablation shows that this metric can be maximized by degenerate standing motions. Unless independent physical-plausibility metrics or quantitative real-robot results are added, the contribution remains a proposal rather than a demonstrated method.

major comments (5)

[§3.3.3, Eq. (5); §4.1 and §4.3, Tables 2–5] The Success Rate used as the headline evidence is exactly the binary reward in Eq. (5). Tables 2–5 evaluate Succ with the same frozen tracking policy π that provides the reward during RL fine-tuning, so the reported gains (e.g., Base 0.43 to RLPF 0.95 on CMU) are by construction a maximization of the evaluation metric. This circularity is not merely formal: Table 5 and Figure 3 show that RLPF-w/o align reaches 0.99 Succ on AMASS while FID collapses to 41.97 and R@1 to 0.07 (Table 6), i.e., the metric can be gamed by degenerate motions. The paper needs an independent physical-validity measure--for example, a tracking policy not used in training, quantitative hardware results, or physics-based plausibility scores--before Succ can be read as evidence of executability.
[§4.3, §5 (real-robot claim)] The claim of successful deployment on a Unitree G1 is supported only by qualitative videos. No quantitative hardware metrics are reported: no tracking error, fall rate, success count, or number of trials. Since the abstract and conclusion assert real-world deployability, this evidence is load-bearing and currently missing.
[§4.2–4.3, Tables 1–3] The baseline set is too narrow to support the stated conclusion. The paper cites physics-aware T2M methods (PhysDiff [Yuan et al. 2023], ReinDiffuse [Han et al. 2025]) and an RL-based motion generator (InstructMotion [Mao et al. 2024]) but does not compare against them. Without these or equivalent physics-aware baselines, "greatly outperforms baseline methods" for physical feasibility is not established.
[§4.4, Tables 4–5] The RLPF-PHC ablation does not resolve the oracle-validity concern. PHC is another AMASS-trained IsaacGym tracking policy, and RLPF-PHC motions are evaluated after transfer to the Exbody2-style student, so the comparison includes a confounded transfer gap and still relies on tracking policies from the same simulated training distribution. In addition, no error bars or multiple-seed results are reported; differences such as 0.92 vs 0.90 on AMASS may not be significant, and the binary nature of Succ makes variance control especially important.
[§5, Limitations] The Limitations paragraph mentions only the frozen-policy generalization constraint. It does not acknowledge that the reward and evaluation share the same oracle or that the paper's own RLPF-w/o align results show this oracle to be gameable. At minimum, the limitation should explicitly state that Succ is a proxy whose validity is not established by the current experiments.

minor comments (7)

[§2] There is a typo: "inscreasing-scale" should be "increasing-scale".
[§3.1] The phrase "interesting a motion tokenizer" should be "integrating a motion tokenizer".
[§4.4] "RLFS" should be "RLPF", and "RLPF-w/o exhibits" should be "RLPF-w/o align exhibits".
[§4.1] The metric group labels are swapped: "High-Level Motion Tracking Metrics" should be "High-Level Motion Generation Metrics", and "Low-Level Motion Generation Metrics" should be "Low-Level Motion Tracking Metrics".
[Eq. (7)] Please clarify whether m denotes the ground-truth motion in the motion alignment verification term; as written, m also appears in Eq. (5) for the retargeted motion, which is confusing.
[§B.2.1] MPJPE is reported in radians (E_mpjpe(rad)); please clarify this convention and confirm that the success thresholds in Section 3.3.3 match those in B.2.1.
[Tables and text] There are formatting issues such as "Base Model0.43" in Table 2 and "Table 4 5 6" in §4.4; these should be cleaned up.

Circularity Check

1 steps flagged · score 6.0 of 10

Primary physical-feasibility metric is the same binary reward optimized by GRPO; RLPF-w/o align demonstrates the reward is gameable.

self definitional [Section 3.3.3, Eq. (5); Section 4.1, Evaluation Metrics]
"R mi tracking = I(Succ(π, mi)) (5) ... Succ(π, mi) is a binary success flag to indicate whether the policy π successfully completes the motion tracking task for mi. Success criteria incorporate motion quality metrics, with failure conditions defined as: positional deviation exceeding threshold ϵ or loss of body balance. ... Success Rate (Succ), which measures whether the humanoid robot follows the reference motion without losing balance (failure threshold: >0.5m average deviation at any timestep) ... Notably, Succ is the primary metric."

RLPF's RL objective is exactly the reported primary metric: GRPO maximizes R_tracking = I(Succ(π, m_i)) from Eq. 5, and Section 4.1 defines Succ as the low-level evaluation metric with the same failure criteria (>0.5m average deviation, loss of balance). Tables 2 and 3 then advertise Succ gains (e.g., 0.48→0.92 on AMASS) as evidence of physical feasibility. This is not an independent check: it measures the very signal being optimized. The paper's own RLPF-w/o align ablation (Table 5, AMASS: Succ 0.99) and Figure 3 show the reward is saturated by degenerate standing motions while all semantic metrics collapse (Table 6: FID 41.97, R@1 0.07), so Succ alone does not certify physical plausibility.

full rationale

The central quantitative claim that RLPF produces physically feasible motions rests on Succ, which is defined as the success flag of the frozen student tracking policy. Eq. 5 uses exactly that flag as the RL reward, so the reported Succ improvements are the training objective measured on held-out motions, not an external physical-validity benchmark. The paper does not validate the tracking policy as a non-gameable oracle; its own RLPF-w/o align ablation shows Succ reaches 0.99 while motions degenerate to standing and semantics collapse, demonstrating that Succ can be maximized without physical plausibility. MuJoCo zero-shot transfer and qualitative real-robot videos provide some external evidence, but the headline tables (Tables 2-5) report the same Succ metric, so the physical-feasibility claim is partially circular. No load-bearing self-citation was found: citations to Exbody2, PHC, H2O, and the authors' earlier large-motion-model work are contextual, not used to justify the reward-evaluation identity. Score 6 reflects that a central prediction reduces by construction while some independent semantic and sim-transfer content remains.

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

No new physical entities are postulated; the tracking policy and alignment encoders are learned artifacts, not new forces, particles, or conserved quantities. The central claim rests instead on the validity of the tracking-policy oracle, simulator transfer, and the semantic encoders, plus hand-chosen reward weights and success thresholds.

free parameters (5)

Tracking success threshold = 0.5 m average deviation; pitch threshold unspecified
Defines both the binary reward R_tracking (Eq. 5) and the primary Succ evaluation metric. Changing the threshold changes the training signal and the reported success rates.
Tracking reward weight = 10
Table 9 sets Reward Weight Tracking to 10; this balances physical feasibility against alignment and KL terms but is chosen by hand without sensitivity analysis.
Alignment reward weight = 2
Table 9 sets Reward Weight Align to 2; this balances semantic alignment against tracking reward and is not swept or justified from data.
KL divergence weight = 1.0
Table 9 sets Weight KL to 1.0, controlling deviation from the reference policy in Eq. 2.
Contrastive margin m = Not reported
Eq. 6 uses a margin threshold m in the contrastive loss for the alignment encoders; the value is not specified in the paper.

assumptions (5)

domain assumption The Exbody2-style student tracking policy trained on AMASS is a valid and transferable oracle for physical feasibility.
Section 3.3.2 trains a teacher-student policy and Section 4.1 uses its success flag as both reward and evaluation. The validity of this oracle is the load-bearing premise for all physical-feasibility claims.
domain assumption IsaacGym dynamics, with zero-shot transfer to MuJoCo, approximate real Unitree G1 physics closely enough to judge executability.
Sections 4.1 and B.1.3 assume simulator tracking performance predicts real-robot tracking, with only qualitative videos as real-world evidence.
domain assumption The contrastive text and motion encoders trained with Eq. 6 capture semantic alignment and are also a fair basis for high-level metrics.
Section 3.4 uses these encoders for RL rewards; Section 4.2 uses pretrained retrieval models for MMDist and R@k. If the same or similar encoders are used, evaluation partially reduces to the training reward.
domain assumption The H2O-style optimization-based retargeting preserves motion semantics while satisfying robot kinematic constraints.
Section 3.3.1 assumes retargeted motions remain semantically equivalent to the human motion after SMPL shape and pose optimization.
standard math GRPO and PPO provide correct policy updates without additional convergence assumptions.
Eq. 2 relies on standard GRPO, and Section 3.3.2 relies on PPO; no new theoretical claims are made.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RL from Physical Feedback: Aligning Large Motion Models with Humanoid Control." pith.science (2026). https://pith.science/paper/4K43SCYC

@misc{pith2026250612769,
  author       = {Pith},
  title        = {Pith review of: RL from Physical Feedback: Aligning Large Motion Models with Humanoid Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4K43SCYC}},
  note         = {Machine review of arXiv:2506.12769}
}

read the original abstract

This paper focuses on a critical challenge in robotics: translating text-driven human motions into executable actions for humanoid robots, enabling efficient and cost-effective learning of new behaviors. While existing text-to-motion generation methods achieve semantic alignment between language and motion, they often produce kinematically or physically infeasible motions unsuitable for real-world deployment. To bridge this sim-to-real gap, we propose Reinforcement Learning from Physical Feedback (RLPF), a novel framework that integrates physics-aware motion evaluation with text-conditioned motion generation. RLPF employs a motion tracking policy to assess feasibility in a physics simulator, generating rewards for fine-tuning the motion generator. Furthermore, RLPF introduces an alignment verification module to preserve semantic fidelity to text instructions. This joint optimization ensures both physical plausibility and instruction alignment. Extensive experiments show that RLPF greatly outperforms baseline methods in generating physically feasible motions while maintaining semantic correspondence with text instruction, enabling successful deployment on real humanoid robots.

Figures

Figures reproduced from arXiv: 2506.12769 by the authors.

**Figure 2.** Overview of RLPF, which consists of three key components: i) Motion Tracking Policy which is pretrained to establish a motion tracking reward to evaluate generated motions; ii) Alignment Verification Module which enhances text-motion semantic consistency while preserving physical plausibility; iii) RL Optimization Framework that jointly optimizes the physical feasibility and semantic alignment of motions generated b… view at source ↗

**Figure 3.** Visualizations of RLPF-w/o align. Since training relies solely on the motion tracking [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Example images of the Isaac Gym and MuJoCo simulators. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

GenTrack: Physical Alignment for Robot-Native Motion Generation and Zero-Shot Humanoid Tracking
cs.RO 2026-08 conditional novelty 6.0 of 10

Online co-training of a text-to-motion generator and a humanoid tracker on simulated G1 improves generator executability and zero-shot tracker coverage beyond static replay or one-way filtering.
PIGEON: VLM-Driven Object Navigation via Points of Interest Selection
cs.RO 2025-11 conditional novelty 6.0 of 10

PIGEON selects sparse Points of Interest for VLM-driven object navigation, achieving state-of-the-art zero-shot ObjectNav results in Habitat and improving a local 7B VLM via distance-based RLVR.
Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos
cs.CV 2025-07 conditional novelty 6.0 of 10

A dexterous VLA pretrained on a 2.5M-instance human hand motion dataset transfers skills to a real robot hand, outperforming baselines in manipulation tasks.
ZeroWBC: Learning Natural Whole-Body Humanoid Interaction from Human Egocentric Data
cs.RO 2026-03 conditional novelty 5.0 of 10

An open-loop generation-then-tracking system maps one egocentric image plus language into Unitree G1 whole-body interactions using only human egocentric motion data.

Reference graph

Works this paper leans on

14 extracted references · 6 canonical work pages · cited by 4 Pith papers

[1]

It reflects the model’s capability to retrieve the most relevant motion

represents the proportion of cases in which the correct motion is ranked as the top match for a given text query. It reflects the model’s capability to retrieve the most relevant motion. R@2 and R@3 follow the same principle, indicating the frequency with which the correct motion appears within the top 2 and top 3 retrieved results, respectively. • Multim...

work page 2023
[2]

Human motion diffusion model.arXiv preprint arXiv:2209.14916,

Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916,

arXiv
[4]

It uses GPU acceleration 12 (a) Isaac Gym example

B.1.1 Isaac Gym Isaac Gym [Makoviychuk et al., 2021] is a high-performance physics simulation environment devel- oped by NVIDIA, designed for robotic reinforcement learning (RL) tasks. It uses GPU acceleration 12 (a) Isaac Gym example

work page 2021
[5]

Momask: Generative masked modeling of 3d human motions

Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910,

work page 1900
[6]

Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470,

Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470,

arXiv
[7]

Since SMPL parameters represent various human body shapes, we first optimize the shape parameter β′ to approximate a humanoid form

Table 7: Comparison of Dataset Quantities Dataset Motion Sequence MotionX 81,082 AMASS 13,145 CMU 5,458 C.2 Motion Retargeting Details As described in line 154 of the main text, we follow the idea of H2O [He et al., 2024] and adopt a two-step optimization-based approach to achieve the retargeting process. Since SMPL parameters represent various human body...

work page 2024
[9]

For computational resources, we use eight A800 GPUs to conduct our experiments. 14 Table 8: Correspondence between Humanoid and Human Body Links Humanoid Links Name Human Body Links Name Pelvis Pelvis Left hip pitch link Left hip Left knee link Left knee Left ankle roll link Left ankle Right hip pitch link Right hip Right knee link Right knee Right ankle ...

work page 2010
[11]

to implement an end-to-end training process, significantly improving the simulation and training speed of complex robotic tasks

Figure 4: Example images of the Isaac Gym and MuJoCo simulators. to implement an end-to-end training process, significantly improving the simulation and training speed of complex robotic tasks. Isaac Gym supports importing standard robot model formats (such as URDF and MJCF), provides Tensor APIs for direct GPU interaction, and can simulate multiple envir...

work page 2012

Show all 14 references

[2007]

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al

[Online]. Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023b. Emanuel Todo...

arXiv
[2011]

Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143,

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,

arXiv
[2012]

Humanoid-gym: Reinforcement learning for humanoid robot with zero-shot sim2real transfer.arXiv preprint arXiv:2404.05695,

Xinyang Gu, Yen-Jen Wang, and Jianyu Chen. Humanoid-gym: Reinforcement learning for humanoid robot with zero-shot sim2real transfer.arXiv preprint arXiv:2404.05695,

arXiv
[2017]

Learning generalizable human motion generator with reinforcement learning.arXiv preprint arXiv:2405.15541,

10 Yunyao Mao, Xiaoyang Liu, Wengang Zhou, Zhenbo Lu, and Houqiang Li. Learning generalizable human motion generator with reinforcement learning.arXiv preprint arXiv:2405.15541,

arXiv
[2020]

Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

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

arXiv
[2023]

Open and efficient foundation language models.Preprint at arXiv

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Open and efficient foundation language models.Preprint at arXiv. https://doi. org/10.48550/arXiv, 2302(3), 202...

doi arXiv

Pith tools

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

[1] [1]

It reflects the model’s capability to retrieve the most relevant motion

represents the proportion of cases in which the correct motion is ranked as the top match for a given text query. It reflects the model’s capability to retrieve the most relevant motion. R@2 and R@3 follow the same principle, indicating the frequency with which the correct motion appears within the top 2 and top 3 retrieved results, respectively. • Multim...

work page 2023

[2] [2]

Human motion diffusion model.arXiv preprint arXiv:2209.14916,

Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model.arXiv preprint arXiv:2209.14916,

arXiv

[3] [4]

It uses GPU acceleration 12 (a) Isaac Gym example

B.1.1 Isaac Gym Isaac Gym [Makoviychuk et al., 2021] is a high-performance physics simulation environment devel- oped by NVIDIA, designed for robotic reinforcement learning (RL) tasks. It uses GPU acceleration 12 (a) Isaac Gym example

work page 2021

[4] [5]

Momask: Generative masked modeling of 3d human motions

Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900–1910,

work page 1900

[5] [6]

Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470,

Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470,

arXiv

[6] [7]

Since SMPL parameters represent various human body shapes, we first optimize the shape parameter β′ to approximate a humanoid form

Table 7: Comparison of Dataset Quantities Dataset Motion Sequence MotionX 81,082 AMASS 13,145 CMU 5,458 C.2 Motion Retargeting Details As described in line 154 of the main text, we follow the idea of H2O [He et al., 2024] and adopt a two-step optimization-based approach to achieve the retargeting process. Since SMPL parameters represent various human body...

work page 2024

[7] [9]

For computational resources, we use eight A800 GPUs to conduct our experiments. 14 Table 8: Correspondence between Humanoid and Human Body Links Humanoid Links Name Human Body Links Name Pelvis Pelvis Left hip pitch link Left hip Left knee link Left knee Left ankle roll link Left ankle Right hip pitch link Right hip Right knee link Right knee Right ankle ...

work page 2010

[8] [11]

to implement an end-to-end training process, significantly improving the simulation and training speed of complex robotic tasks

Figure 4: Example images of the Isaac Gym and MuJoCo simulators. to implement an end-to-end training process, significantly improving the simulation and training speed of complex robotic tasks. Isaac Gym supports importing standard robot model formats (such as URDF and MJCF), provides Tensor APIs for direct GPU interaction, and can simulate multiple envir...

work page 2012