Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Multi-Loco: Unifying Multi-Embodiment Legged Locomotion via Reinforcement Learning Augmented Diffusion

T0 review · 3 major / 4 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 Multi-Loco claims that one shared policy, a frozen cross-embodiment diffusion prior plus a shared residual reinforcement-learning policy, can control four morphologically different legged robots and outperform per-robot PPO baselines by…

desk verdict Mask convention in the core objective is self-contradictory (Eq. 2 vs. Algorithm 1 and A.2), and without code the headline results are unverifiable; the architecture itself is worth a referee's time. read the letter →

arxiv 2506.11470 v1 pith:D34DVYO2 submitted 2025-06-13 cs.RO

Shunpeng Yang , Zhen Fu , Zhefeng Cao , Guo Junde , Patrick Wensing , Wei Zhang , Hua Chen This is my paper · ORCID
classification cs.RO
keywords multi-embodimentlocomotiondiffusionpolicyresidualreinforcementlearningmaskedscorematchingleggedrobotssim2realtransferPPOcross-embodiment
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

Multi-Loco proposes that a single shared policy can control four morphologically different legged robots by splitting the job in two: a frozen diffusion model, trained offline on padded, cross-embodiment data, supplies a general action prior, and a lightweight residual policy trained with reinforcement learning refines that prior online. The paper reports that the combined policy outperforms per-robot PPO baselines by 10.35% in average return, with gains up to 13.57% on a wheeled biped, and shows zero-shot transfer to physical robots and to an unseen quadruped morphology. The intended significance is that locomotion skills acquired by one body design can be transferred to others, reducing the need to retrain every platform from scratch.

What carries the argument

The load-bearing object is the masked denoising score-matching objective of Equation (2), which is a training loss that applies the diffusion reconstruction error only on the valid dimensions of a zero-padded action vector, using a binary mask $b$ with 1 for real action dimensions and 0 for padding. This lets a single EDM/DiT denoiser learn the action distribution of every embodiment from one shared dataset. The second component is a residual PPO policy with one shared actor and one critic per robot; the actor outputs an additive correction $\Delta a$ to the diffusion prior, and the per-robot critics stabilize optimization across conflicting reward scales. The mask defines which data the prior learns from, and the residual defines which task-specific corrections survive online.

What would settle it

Compare the mask operations in the implementation with the two conventions in the paper: if the training loss is $((1-b)\odot(D_\theta(\cdot)-a_0))^2$ and inference applies $(1-b)\cdot a_i$, as Algorithm 1 and Section A.2 state, then valid action dimensions are never trained and are dropped at deployment; re-running with the Equation (2) convention (loss weighted by $b$, inference zeroing only padded dimensions) would either reproduce the reported 10.35% average gain or show the mask flip was responsible.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that morphology-agnostic representation is enough to unify legged locomotion across a point-foot biped, a wheeled biped, a humanoid, and a quadruped, without explicit morphology descriptors, joint information, or per-robot encoders. This is achieved by zero-padding all observations and actions into common spaces and training one EDM denoiser with masked denoising score matching, so the diffusion loss acts only on each robot's valid action dimensions. The denoiser's output is then added to a residual action predicted by a single PPO actor shared across all embodiments, while a separate critic per robot provides the advantage estimates. The authors report that the full method reaches 113.57% of the baseline's average return, that cross-robot diffusion training beats single-robot diffusion training by 17.96%, and that the same policy runs in real time at 50 Hz on hardware.

Load-bearing premise

The load-bearing premise is that the binary mask used in training and inference matches Equation (2), where the mask marks valid action dimensions; the paper's own Algorithm 1 and Section A.2 apply the complement mask $(1-b)$, so if those describe the implementation, the diffusion model would train on padded dimensions only and would zero out the real actions at deployment.

Editorial extensions

If this is right

  • A single shared policy can replace per-robot control stacks on four different legged morphologies, so locomotion data collected on any of them can contribute to one controller.
  • Cross-embodiment diffusion training transfers terrain skills: the wheeled biped acquires leg-lifting behavior on discrete terrain that its own single-robot diffusion model and RL baseline never produced.
  • The diffusion prior alone reaches 87.78% of the baseline return, and the residual RL lifts this to 113.57%, showing the generative prior and the online policy are complementary rather than redundant.
  • The same unified policy transfers zero-shot to a previously unseen quadruped morphology and to physical hardware on grass, slopes, stairs, and gravel paths.

Reading between the lines

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

  • Editorial inference: the padding-plus-mask recipe could generalize beyond legs to any control task with heterogeneous action spaces, such as manipulators with different numbers of fingers, provided the mask convention between training and inference is kept consistent.
  • Editorial inference: the paper's finding that reducing humanoid data degrades wheeled-biped performance suggests high-dimensional whole-body datasets act as a general prior for lower-dimensional locomotion; one could test whether adding more humanoid data monotonically improves wheeled-biped terrain metrics.
  • Editorial inference: a clean ablation that freezes or removes the residual policy at deployment would isolate how much of the reported sim2real robustness comes from the diffusion prior versus the RL refinement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Multi-Loco, a framework for multi-embodiment legged locomotion that combines a frozen, cross-embodiment diffusion prior (trained with masked denoising score matching) with a shared residual policy trained by multi-critic PPO. The action space is unified across four robot types by zero-padding, and the claimed contribution is that the hybrid policy outperforms per-robot PPO baselines by 10.35% average return, with gains up to 13.57% on a wheeled biped, while also transferring zero-shot to a Unitree Go2 and to real hardware. The manuscript reports simulation and real-world experiments, hyperparameter tables, and several ablations of dataset size, sampling steps, and dataset composition.

Significance. If the results are verified, the paper would make a useful contribution to cross-embodiment locomotion: it demonstrates a concrete architecture for combining a morphology-agnostic diffusion prior with a shared residual RL policy, and it reports real-world deployment on four distinct platforms plus zero-shot transfer to an unseen quadruped. The authors provide substantial experimental detail, including reward tables, network configurations, and an appendix with ablations. However, the central quantitative claims currently lack the necessary supporting evidence: the mask-convention inconsistency directly affects the correctness of the diffusion prior, the headline return numbers are not backed by a numeric table with errors, and the diffusion prior is trained on rollouts from the very baseline it is compared against. These issues are fixable, but they are load-bearing rather than cosmetic.

major comments (3)
  1. [Section 2.2 / Algorithm 1 / Section A.2] The mask convention is self-contradictory in a way that directly affects the validity of the diffusion prior. Section 2.1 defines b with 1 for valid action dimensions and 0 for padded dimensions, and Eq. (2) multiplies the denoising loss by b, so only valid dimensions are supervised. In contrast, Algorithm 1 line 26 and Section A.2 compute L = ((1-b)*(D_theta(...)-a0))^2, and Algorithm 1 line 38 sets a'_i = (1-b)*a_i during inference. Under the Algorithm 1 convention, the network is trained only on padded dimensions and the valid dimensions are zeroed at inference, the exact complement of Eq. (2). Since no code or data is released, a reader cannot determine which convention the reported experiments actually used; this is load-bearing because the central 10.35% and 13.57% claims depend on the frozen diffusion prior supplying useful cross-embodiment actions. The authors must correct the inconsistency and, ideally, release code or detailed pseudocode that matches the implementation.
  2. [Section 3.2, Table 2, Figure 3] The headline average return improvement is not supported by a numeric table. Table 2 reports only MEL, LVT, and AVT, while the claimed 10.35% average improvement and per-robot gains (12.49%, 13.57%, 4.38%, 10.97%) appear only in the caption of Figure 3, which is a bar chart without error bars or numerical values. For the quadruped, Table 2 shows CR-DP+RA slightly underperforming the RL baseline on all three reported metrics (MEL 19.60 vs. 19.79, LVT 5.42 vs. 5.45, AVT 4.32 vs. 4.34), which is difficult to reconcile with a claimed 4.38% return gain unless average return captures a different quantity. Please provide a table of Average Return with means and standard deviations across seeds for every method and embodiment.
  3. [Section 3.2, Evaluation Setup] The diffusion prior is not an external benchmark because it is distilled from rollouts of the same RL baselines it is later compared against. The text states: 'Our evaluation protocol employs policies trained via baseline RL to generate rollout data for diffusion policy (DP) training.' This makes the SR-DP/CR-DP comparisons informative only about behavioral cloning from the baseline, not about whether the diffusion prior provides knowledge beyond the baseline. The final residual policy is trained with environment rewards, which gives some independent grounding, but the contribution of the diffusion prior itself is not isolated: there is no ablation of a shared residual policy without the diffusion prior (e.g., a Gaussian-prior residual PPO with the same multi-critic architecture and reward weights). Adding such an ablation would clarify whether the reported gains come from the diffusion prior, the residual RL, or the larger shared architecture.
minor comments (4)
  1. [Section A.2] The sentence 'we zero-pad observation and action vectors to match the largest dimension across all robot morphologies (observations: 68D, actions: 20D) and the dimension configurations of .' is incomplete; it ends with 'of .' and appears to reference a table or equation that is not included.
  2. [Figure 3 caption] The caption labels the full method as 'SR-DP+RA (Ours)', but the method introduced in Section 3.2 is 'CR-DP with Residual Adaptation (CR-DP+RA)'. This likely a typo and should be corrected to avoid confusion.
  3. [Algorithm 1 / Table 3] There are inconsistent hyperparameter values: Algorithm 1 line 5 specifies a batch size of 512 in the initialization block, while Table 3 lists 'Batch Size 8192'. Similarly, Algorithm 1's training procedure is written as 'Epochs=400K' while Table 3 reports 'Number of Epochs 500'. Please align these values.
  4. [Section 2.3, Eq. (4)] The actor loss in Eq. (4) is written as a weighted log-probability objective, but it omits the importance-sampling ratio and clipping that characterize the PPO update described in the text. A reader implementing from Eq. (4) would not reproduce standard PPO; please either present the full clipped surrogate loss or explain the simplification.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the reported gains are empirical measurements and the residual RL stage provides independent grounding.

full rationale

Multi-Loco does not present a derivation chain in which an output is defined as its input. The central comparison (CR-DP+RA vs RL baseline) is empirical: although the diffusion prior is trained on rollouts generated by the baseline RL policies (Section 3.2: 'Our evaluation protocol employs policies trained via baseline RL to generate rollout data for diffusion policy (DP) training'), the final action is a_prior + Delta and Delta is optimized with environment rewards, so the reported 10.35% improvement is not equal to the baseline by construction. This training-data overlap is a benchmarking confound, not a circular reduction. Likewise, CR-DP vs SR-DP is confounded by data volume (four datasets vs one), not by definition. The paper contains self-citations (e.g., [9], [28], [29]) but none is load-bearing for the main claim. A separate, non-circular reproducibility problem is the mask-convention contradiction: Eq. (2) multiplies the loss by b (valid dimensions), while Algorithm 1 line 26 and Section A.2 use (1-b) and inference zeros valid dimensions; this makes the experimental results unverifiable from the manuscript alone, but it is an internal inconsistency, not a reduction of the prediction to its inputs. No uniqueness arguments are imported from prior author work. Therefore no circularity is established.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on the EDM/DiT diffusion prior, zero-padding alignment, the RL residual with per-robot critics, and the quality of baseline-generated datasets. None of these is verified by external benchmarks, released code, or public data, and several involve author-selected coefficients, so the ledger lists the main parameters and assumptions the reader is asked to accept.

free parameters (5)
  • Residual penalty coefficient alpha = 0.2 (PPO Residual Coeff)
    Hand-chosen coefficient in the residual penalty rd(delta a_t)=alpha*||delta a_t||_1, introduced in Section 2.3 and Table 6; it controls how much the RL correction may deviate from the diffusion prior.
  • Per-robot reward weights = Tables 8 and 9, e.g. velocity gains 4.0/2.0/2.0/6.0 across robots
    Each robot uses different coefficients for velocity tracking, base height, orientation, regularization, and contact penalties; these are author-selected and partly drive the reported returns.
  • Desired base height for a1 and go2 = Not numerically reported, Appendix A.6
    The paper explains a large Average Return gap between a1 and go2 by differing desired base height parameters and adjusts the comparison by subtracting that reward, so the zero-shot claim depends on an unreported parameter.
  • Diffusion sampling steps = 5
    Selected from an ablation in Appendix A.3; fewer steps reduce inference time but may degrade action quality, so the deployed results depend on this choice.
  • Dataset size cap per robot = 2,048,000 samples
    Maximum samples per robot in Table 3; the CR-DP versus SR-DP comparison mixes dataset volume with cross-embodiment content, so this cap is important for the emergence claim.
assumptions (4)
  • standard math EDM and score-matching theory correctly model the multimodal action distribution, and the 5-step Euler ODE solver produces accurate actions.
    Invoked in Section 2.2 and Appendix A; relies on Song et al. and Karras et al. without re-derivation.
  • domain assumption Rollouts from per-robot RL baseline policies are sufficient and representative training data for the diffusion prior.
    Section 3.2 trains SR-DP and CR-DP on baseline rollouts; if the data is narrow or suboptimal, the prior inherits those limitations.
  • domain assumption Zero-padded fixed-size observation and action vectors preserve enough information for one shared network to infer each robot's embodiment and dynamics without explicit morphology descriptors.
    Section 2.1 and the unified observation and action design rely on this; no ablation varies the alignment scheme.
  • ad hoc to paper A multi-critic PPO with one shared actor resolves optimization conflicts across embodiments better than alternatives.
    Equation (4) and Table 7 specify this architecture; the paper does not compare against a single-critic or per-robot actor baseline, so the necessity of this design is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Loco: Unifying Multi-Embodiment Legged Locomotion via Reinforcement Learning Augmented Diffusion." pith.science (2026). https://pith.science/paper/D34DVYO2

@misc{pith2026250611470,
  author       = {Pith},
  title        = {Pith review of: Multi-Loco: Unifying Multi-Embodiment Legged Locomotion via Reinforcement Learning Augmented Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D34DVYO2}},
  note         = {Machine review of arXiv:2506.11470}
}
read the original abstract

Generalizing locomotion policies across diverse legged robots with varying morphologies is a key challenge due to differences in observation/action dimensions and system dynamics. In this work, we propose Multi-Loco, a novel unified framework combining a morphology-agnostic generative diffusion model with a lightweight residual policy optimized via reinforcement learning (RL). The diffusion model captures morphology-invariant locomotion patterns from diverse cross-embodiment datasets, improving generalization and robustness. The residual policy is shared across all embodiments and refines the actions generated by the diffusion model, enhancing task-aware performance and robustness for real-world deployment. We evaluated our method with a rich library of four legged robots in both simulation and real-world experiments. Compared to a standard RL framework with PPO, our approach -- replacing the Gaussian policy with a diffusion model and residual term -- achieves a 10.35% average return improvement, with gains up to 13.57% in wheeled-biped locomotion tasks. These results highlight the benefits of cross-embodiment data and composite generative architectures in learning robust, generalized locomotion skills.

Figures

Figures reproduced from arXiv: 2506.11470 by the authors.

Figure 1
Figure 1. Deployment of the reinforcement learning augmented diffusion policy on four platforms (biped, wheeled biped, humanoid and quadruped). The experimental results demonstrate that the unified policy can effectively control the robots across various types of uneven terrain, including grass, slopes, stairs, and gravel paths. These results highlight the policy’s robustness and exceptional control capabilities. Abstract: Ge… view at source ↗
Figure 2
Figure 2. Overview of the Multi-Loco framework. Multi-robot datasets are preprocessed via zero-padding and normalization to align observation and action spaces across embodiments. A shared diffusion model is trained offline using masked denoising score matching. At inference time, the diffusion model generates action priors, which are refined by a residual policy trained via multi-critic PPO. Each critic specializes in one ro… view at source ↗
Figure 3
Figure 3. (a) Comparative performance analysis of four robot morphologies (biped, humanoid, wheeled, quadruped) in terrain negotiation tasks. SR-DP+RA achieves 10.35% average improvement over RL baseline (12.49% biped, 13.57% wheeled-biped, 4.38% quadruped, 10.97% humanoid), while CR-DP surpasses SR￾DP by 17.96% (17.81% biped, 7.52% wheeled-biped, 20.47% quadruped, 26.02% humanoid). (b) Humanoid locomotion demonstrates zero-s… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Terrain traversal performance of wheeled-biped robots under different training setups. (a) Performance of the diffu￾sion model trained using offline datasets only (SR-DP-Wheeled vs. CR-DP) on terrains seen during training. (b) Comparison between PPO-trained baseline an…
Figure 5
Figure 5. Figure 5: Neural network structure of DiT which is used to fit [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Comparison with Different Sampling Steps in EDM [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Biped Statistics: Comparison with EDMs Trained with Different Datasets Size. (The [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Impact of dataset composition ratios on diffusion policy training. The baseline normal ratio [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Zero-Shot Transfer to Unitree Go2 To evaluate generalization beyond training morphologies, we directly deploy the unified policy on Unitree Go2 without any finetun￾ing. Despite being unseen during training, the policy achieves stable forward locomo￾tion, demonstrating …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 11 canonical work pages

  1. [1]

    Gu, Y .-J

    X. Gu, Y .-J. Wang, X. Zhu, C. Shi, Y . Guo, Y . Liu, and J. Chen. Advancing humanoid loco- motion: Mastering challenging terrains with denoising world model learning.arXiv preprint arXiv:2408.14472, 2024

    arXiv 2024

  2. [2]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learn- ing agile and dynamic motor skills for legged robots.Science Robotics, 4(26):eaau5872, 2019

    work page 2019

  3. [3]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

    2020

  4. [4]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning robust percep- tive locomotion for quadrupedal robots in the wild.Science robotics, 7(62):eabk2822, 2022

    work page 2022

  5. [5]

    J. Yang, D. Sadigh, and C. Finn. Polybot: Training one policy across robots while embracing variability.arXiv preprint arXiv:2307.03719, 2023

    arXiv 2023

  6. [6]

    M. Xu, Z. Xu, C. Chi, M. Veloso, and S. Song. Xskill: Cross embodiment skill discovery. In Conference on Robot Learning, pages 3536–3555. PMLR, 2023

    work page 2023

  7. [7]

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy.arXiv preprint arXiv:2405.12213, 2024

    arXiv 2024

  8. [8]

    Doshi, H

    R. Doshi, H. Walke, O. Mees, S. Dasari, and S. Levine. Scaling cross-embodied learn- ing: One policy for manipulation, navigation, locomotion and aviation.arXiv preprint arXiv:2408.11812, 2024

    arXiv 2024

Show all 33 references
  1. [9]

    H. Wang, H. Luo, W. Zhang, and H. Chen. Cts: Concurrent teacher-student reinforcement learning for legged locomotion.IEEE Robotics and Automation Letters, 2024

    2024

  2. [10]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, and D. Pathak. Extreme parkour with legged robots. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 11443–11450. IEEE, 2024. 9

    2024

  3. [11]

    Hoeller, N

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter. Anymal parkour: Learning agile navigation for quadrupedal robots.Science Robotics, 9(88):eadi7566, 2024

    2024

  4. [12]

    J. Lee, M. Bjelonic, A. Reske, L. Wellhausen, T. Miki, and M. Hutter. Learning robust au- tonomous navigation and locomotion for wheeled-legged robots.Science Robotics, 9(89): eadi9641, 2024

    2024

  5. [13]

    T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, and G. Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning.arXiv preprint arXiv:2406.08858, 2024

    2024 arXiv

  6. [14]

    T. He, W. Xiao, T. Lin, Z. Luo, Z. Xu, Z. Jiang, J. Kautz, C. Liu, G. Shi, X. Wang, et al. Hover: Versatile neural whole-body controller for humanoid robots.arXiv preprint arXiv:2410.21229, 2024

    2024 arXiv

  7. [15]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based gen- erative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  8. [16]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  9. [17]

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in Neural Information Processing Systems, 35:5775–5787, 2022

    2022

  10. [18]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    2022

  11. [19]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, page 02783649241273668, 2023

    2023

  12. [20]

    Huang, Y

    X. Huang, Y . Chi, R. Wang, Z. Li, X. B. Peng, S. Shao, B. Nikolic, and K. Sreenath. Diffuse- loco: Real-time legged locomotion control with diffusion from offline datasets.arXiv preprint arXiv:2404.19264, 2024

    2024 arXiv

  13. [21]

    G. Feng, H. Zhang, Z. Li, X. B. Peng, B. Basireddy, L. Yue, Z. Song, L. Yang, Y . Liu, K. Sreenath, et al. Genloco: Generalized locomotion controllers for quadrupedal robots. In Conference on Robot Learning, pages 1893–1903. PMLR, 2023

    1903

  14. [22]

    Shafiee, G

    M. Shafiee, G. Bellegarda, and A. Ijspeert. Manyquadrupeds: Learning a single locomotion policy for diverse quadruped robots. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 3471–3477. IEEE, 2024

    2024

  15. [23]

    Z. Luo, Y . Dong, X. Li, R. Huang, Z. Shu, E. Xiao, and P. Lu. Moral: Learning morphologically adaptive locomotion controller for quadrupedal robots on challenging terrains.IEEE Robotics and Automation Letters, 2024

    2024

  16. [24]

    Bohlinger, G

    N. Bohlinger, G. Czechmanowski, M. Krupka, P. Kicki, K. Walas, J. Peters, and D. Tateo. One policy to run them all: an end-to-end learning approach to multi-embodiment locomotion. arXiv preprint arXiv:2409.06366, 2024

    2024

  17. [25]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

    2023

  18. [26]

    Mysore, G

    S. Mysore, G. Cheng, Y . Zhao, K. Saenko, and M. Wu. Multi-critic actor learning: Teaching rl policies to act with style. InInternational Conference on Learning Representations, 2022. 10

    2022

  19. [27]

    P. Xu, X. Shang, V . Zordan, and I. Karamouzas. Composite motion learning with task control. ACM Transactions on Graphics (TOG), 42(4):1–16, 2023

    2023

  20. [28]

    Huang, J

    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.arXiv preprint arXiv:2502.08378, 2025

    2025 arXiv

  21. [29]

    H. Wang, Z. Wang, J. Ren, Q. Ben, T. Huang, W. Zhang, and J. Pang. Beamdojo: Learning agile humanoid locomotion on sparse footholds.arXiv preprint arXiv:2502.10363, 2025

    2025 arXiv

  22. [30]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. InConference on Robot Learning, pages 91–100. PMLR, 2022

    2022

  23. [31]

    Ross and D

    S. Ross and D. Bagnell. Efficient reductions for imitation learning. InProceedings of the thir- teenth international conference on artificial intelligence and statistics, pages 661–668. JMLR Workshop and Conference Proceedings, 2010

    2010

  24. [32]

    D. Picard. Torch. manual seed (3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision.arXiv preprint arXiv:2109.08203, 2021

    2021 arXiv

  25. [2024]

    elu", "elu

    URLhttps://arxiv.org/abs/2406.09509. 11 Appendix Experiment Videos We conducted comprehensive real-world evaluations of our framework across four distinct legged robotic platforms. For detailed empirical validation, we encourage readers to view the supplemen- tary video. As de...

    arXiv

Pith tools

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