Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Motion Priors Reimagined: Adapting Flat-Terrain Skills for Complex Quadruped Mobility

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A frozen flat-ground motion prior plus learned joint residuals transfers animal gaits to rough terrain.

desk verdict A genuinely working hierarchical system with an unproven attribution claim: the flat-ground prior's causal contribution is never isolated from the residual penalty and architecture. read the letter →

arxiv 2505.16084 v2 pith:CHB2UNID submitted 2025-05-21 cs.RO

classification cs.RO
keywords motionimitationpriorhierarchicalreinforcementlearningjointresidualsquadrupedlocomotionlocalnavigationsim-to-realtransferterrainadaptation
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 sets out to show that a robot does not need motion data from rough terrain to move well on rough terrain. It pre-trains a low-level policy to imitate walking, pacing, and cantering from flat-ground animal motion capture, freezes it, and then trains a high-level policy that selects among those gaits and adds small joint corrections, called residuals. The claim is that these residuals let a quadruped climb stairs, traverse slopes, and step around obstacles while keeping the natural animal-like gait, and that the motion prior regularizes learning so fewer hand-tuned style rewards are needed. The paper reports simulation success rates of 75–95 percent across terrain types and real-world demonstration on the ANYmal-D robot.

What carries the argument

The machinery has three parts. First, a Fourier latent dynamics encoder compresses retargeted animal motion clips into latent variables that capture frequency, amplitude, offset, and phase, and a low-level policy is trained on flat ground to imitate motions conditioned on those latents. Second, the frozen low-level policy is exposed to a high-level teacher policy that outputs a 16-dimensional latent command $z_t$ and 12-dimensional joint residuals $a^{\mathrm{res}}_t$; a penalty $r_{\mathrm{res}} = w_{\mathrm{res}}\sum_{i=1}^{12}(a^{\mathrm{res}}_{t,i})^2$ is the main lever controlling how far the behavior may depart from the prior. Third, teacher-student distillation with a GRU belief encoder transfers the teacher's actions to a student that sees only noisy proprioceptive and exteroceptive observations.

What would settle it

Deploy the distilled policy on terrains explicitly excluded from training—gaps, stepping stones, or overhanging obstacles—and measure goal-reaching success from 100 randomized starts; if success falls to near chance, the claim that flat-ground priors generalize to complex, rugged terrain is bounded to the specific terrain distribution used in training.

Watch

Extended reading notes

Core claim

The core discovery is that the flat-terrain-only motion prior is not a limitation once a high-level policy is allowed to emit latent gait commands plus bounded joint residuals. With the low-level policy frozen, the high-level teacher learns to adjust the gait to the terrain; the residual penalty weight becomes a single dial that trades off style preservation against adaptability. Distilling this teacher into a student policy with a recurrent belief encoder transfers the behavior to hardware, where the robot keeps the animal-like gait while performing goal-directed local navigation.

Load-bearing premise

The whole argument rests on the assumption that the simulated sensor and physics profiles used in teacher-student distillation are close enough to the real ANYmal-D robot and its LiDAR that the simulated behavior transfers to hardware; the paper does not disclose the noise models, domain-randomization ranges, or terrain physics parameters used to close that gap.

Editorial extensions

If this is right

  • A single frozen low-level policy can serve as a shared motion prior for multiple downstream tasks, since the high-level policy only emits latent commands and residuals.
  • Success on rugged terrain becomes tunable with one scalar: residual penalty weight; too high a penalty caps adaptation, too low a penalty lets gaits drift off the prior.
  • Task policies trained on top of motion priors need far fewer reward terms; with minimal reward terms the same priors still produce a cantering gait that navigates and avoids obstacles.
  • The distilled policy performs goal-directed local navigation without a separate waypoint planner, using only onboard exteroception.
  • Motion priors reduce cost of transport compared with a from-scratch policy under equivalent rewards, because the natural gait avoids wasteful vertical motion.

Reading between the lines

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

  • Beyond the paper: the architecture suggests a skill-library view of motion priors; adding jump, crawl, or turning latents and sampling them during task training could widen the terrain envelope without retraining the low-level policy.
  • Beyond the paper: the residual variance plots imply a per-joint residual diagnostic could be used online to estimate terrain difficulty, but the paper does not use residuals that way.
  • Beyond the paper: replacing the imitation objective with an adversarial style reward on the same flat-ground data would test whether the result is specific to latent priors; the paper's baseline only removes the prior entirely.
  • Beyond the paper: the cost-of-transport comparison is retrospective, so an explicit energy objective combined with a mild residual penalty could likely lower energy use further.
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

4 major / 5 minor

Summary. The paper proposes a hierarchical RL framework for quadruped locomotion. A low-level policy is pre-trained on flat-ground retargeted animal mocap using an FLD latent representation, then frozen and used as a motion prior. A high-level, goal-conditioned teacher policy is trained on rough terrains; it outputs a latent command to the low-level policy plus 12D joint residuals, with a penalty on the residuals. The teacher is distilled into a GRU-based student policy with noisy observations. Experiments in Isaac Gym cover stairs, slopes, random boxes, and obstacles; the authors report success rates, residual-variance analyses, and a from-scratch baseline comparison. Hardware experiments on ANYmal-D demonstrate goal-directed navigation with an animal-like gait on stairs and random blocks. The central claim is that flat-terrain motion priors plus residuals enable natural, regularized locomotion and local navigation on complex terrains without requiring rough-terrain demonstration data.

Significance. If the central claim is established, this is a useful contribution: it shows that raw flat-ground mocap priors can be adapted by a residual high-level policy to rough terrain, avoiding the need to collect demonstration data in the target environment and reducing reward tuning for motion style. The paper has clear strengths: multiple gaits, multiple terrain types, a residual-penalty ablation, an energy-efficiency comparison, and a real-robot ANYmal-D demonstration. The authors also state limitations candidly, including mode collapse. The main weakness is that the attribution of the natural gait to the motion prior is not isolated by the current baseline comparison, and the regularization metric is partly circular because it uses the same FLD representation used for training. These issues are fixable with additional experiments, so the result is promising but not yet fully established.

major comments (4)
  1. [Sec. 4.2.4, Eq. (3)] The comparison against the from-scratch baseline does not isolate the contribution of the motion prior. The baseline has a different action space (12D direct joint actions instead of latent commands plus residuals) and does not include the residual penalty of Eq. (3), so the observed difference in gait could be caused by action parameterization, model capacity, or regularization strength rather than by the flat-ground pretraining. Please add a matched baseline with the same hierarchical architecture, the same latent-command/residual action interface, and the same residual penalty, but with an untrained or randomly initialized low-level policy. Report success rates, joint tracking error against the raw retargeted mocap references, footfall patterns, and CoT for that baseline. The limitation stated in Sec. 6 that the high-level policy can mode-collapse onto a single gait and use residuals only for adaptation makes this matched ablation especially necessary.
  2. [Sec. 4.2.3, Fig. 7] The motion-regularization metric is computed as the deviation of actual joint positions from FLD decoder reconstructions, i.e., from the same representation used to train the low-level policy. This makes the regularization claim partially circular: a low deviation from FLD reconstructions may reflect that the FLD latent space is easy to stay near, not that the policy is faithfully reproducing the animal motion. Please also evaluate against the raw retargeted mocap reference directly, or against an independent style metric such as footfall phase, joint-acceleration spectrum, or base vertical velocity, and report aggregate values over all joints rather than only a representative joint.
  3. [Table 8, Fig. 6] Success rates are reported as x/100 from a single evaluation run, without variance over training seeds. The residual-penalty ablation (Fig. 6) and the baseline comparison (Sec. 4.2.4) are load-bearing for the paper's claims, so single-run numbers are not sufficient. Please report mean and standard deviation over at least three to five training seeds, and describe how the 100 trials per seed are randomized (initial pose, goal direction, terrain instance).
  4. [Sec. 3.4, Sec. 4.3] The sim-to-real transfer section does not disclose the sensor noise models, domain randomization ranges (dynamics, friction, mass, payload), or terrain randomization used for student training. Without these details, the hardware results in Sec. 4.3 cannot be reproduced or generalized beyond the specific test course. Please add a table listing the noise parameters and randomization ranges, and, if available, a robustness check over randomization seeds.
minor comments (5)
  1. [Sec. 3.2] 'retargetted' should be 'retargeted'; the sentence 'These modifications may help reduce performance degradation from FLD decoder reconstruction errors' is vague about whether the decoder is used at all during low-level training, so please state explicitly which components of the FLD model are frozen and which are discarded.
  2. [Sec. 4.2.2] The statement 'peaks with continuously increasing height can be observed in Fig. 4' is based on visual inspection; please specify the difficulty levels being compared and provide quantitative peak or variance values for the representative joints.
  3. [Eq. (1)] The condition 't > T - Tr' is unclear because T is defined as the interval between successive position commands; please clarify the timing logic and define T and Tr with explicit units.
  4. [Table 2] The 'Stand Still' and 'Stand Pose' rewards are not described in the main text; please clarify whether they are active in all high-level experiments and give their weights in the table caption.
  5. [Sec. 8.4, Fig. 12] The CoT comparison has no error bars or trial counts; please add the number of episodes and standard deviations so the energy-efficiency claim can be evaluated quantitatively.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: external benchmarks ground the central claim, while the FLD-based style metric is only a partially self-referential diagnostic.

full rationale

The paper's central derivation—pre-training a frozen flat-ground imitation policy as a motion prior and adding a high-level residual policy for rough-terrain goal reaching—is evaluated against external benchmarks (success rates on stairs, slopes, boxes, and high obstacles in Table 8, plus real-world ANYmal-D deployment), so the main claim is not equivalent to its inputs by construction. The comparative motion-regularization claim in Sec. 4.2.4 is weakened by a confounded baseline (12D direct joint actions, no residual penalty, no latent command interface), but a confounded ablation is an experimental attribution gap rather than circularity: the observed jumping gait of the baseline is not logically forced by the equations. Fig. 7's style metric uses the FLD decoder, which belongs to the same representation family as the low-level prior, making it a partially self-referential diagnostic; however, the decoder is a fixed function trained on the original animal mocap and is used only as a proxy, not as the definition of the paper's central result. Self-citations such as [18] are method citations and are not load-bearing proofs; no uniqueness theorem or fitted parameter is renamed as a prediction. The paper's own Sec. 6 limitation about mode collapse further underscores that the prior's causal contribution is not fully isolated, but that is a limitation, not a circular step. Overall, no load-bearing step reduces to its own input by construction.

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

The framework relies on several domain assumptions about transferability (flat-ground priors to rough terrain, simulation to hardware, latent representation sufficiency) and on hand-tuned reward weights including the residual penalty. It introduces no new physical or ontological entities; the latent commands and joint residuals are learned control variables, not newly postulated entities.

free parameters (4)
  • Joint residual penalty weight wres = -0.1 (main configuration); swept 0 to -100
    Balances task completion against adherence to the motion prior; chosen by evaluating success rate and joint error (Fig 6, Fig 7).
  • Low-level imitation reward weights = various (e.g., joint position tracking 1.4, linear velocity tracking 2.0, angular velocity tracking 0.8)
    Hand-tuned in the low-level imitation reward (Table 1) to achieve good tracking on flat ground; the resulting policy is the motion prior that the whole framework depends on.
  • Goal reach distance threshold = 2 m (activation threshold for rreach in Eq. 1)
    Sparse reward threshold chosen for the navigation task; affects the curriculum of task learning.
  • High-level task reward weights = 15 (position tracking) and 5 (heading velocity)
    Chosen to balance goal-reaching with smooth progression; these weights shape the high-level teacher policy (Table 2).
assumptions (4)
  • domain assumption Flat-ground mocap data retargeted to ANYmal-D provides a valid and sufficient motion prior for rough terrain when augmented with residuals.
    The entire framework depends on this transfer of flat-ground skills; Sec 3.2-3.3.
  • domain assumption Isaac Gym simulation with the described terrain and sensor models is faithful enough that the distilled student policy transfers to the real ANYmal-D.
    Sim-to-real transfer is claimed via teacher-student distillation (Sec 3.4) and validated only qualitatively on hardware (Sec 4.3).
  • domain assumption The FLD encoder's latent representation (frequency, amplitude, offset, phase) captures the essential motion characteristics for commanding gaits.
    The low-level policy relies on this representation; the paper replaces the FLD decoder with ground-truth trajectories, implying the encoder alone is sufficient (Sec 3.2).
  • ad hoc to paper Residual joint corrections are sufficient to adapt the frozen low-level policy to novel terrains.
    This is the paper's core design choice; there is no fine-tuning of the low-level policy (Sec 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Motion Priors Reimagined: Adapting Flat-Terrain Skills for Complex Quadruped Mobility." pith.science (2026). https://pith.science/paper/CHB2UNID

@misc{pith2026250516084,
  author       = {Pith},
  title        = {Pith review of: Motion Priors Reimagined: Adapting Flat-Terrain Skills for Complex Quadruped Mobility},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CHB2UNID}},
  note         = {Machine review of arXiv:2505.16084}
}
read the original abstract

Reinforcement learning (RL)-based motion imitation methods trained on demonstration data can effectively learn natural and expressive motions with minimal reward engineering but often struggle to generalize to novel environments. We address this by proposing a hierarchical RL framework in which a low-level policy is first pre-trained to imitate animal motions on flat ground, thereby establishing motion priors. A subsequent high-level, goal-conditioned policy then builds on these priors, learning residual corrections that enable perceptive locomotion, local obstacle avoidance, and goal-directed navigation across diverse and rugged terrains. Simulation experiments illustrate the effectiveness of learned residuals in adapting to progressively challenging uneven terrains while still preserving the locomotion characteristics provided by the motion priors. Furthermore, our results demonstrate improvements in motion regularization over baseline models trained without motion priors under similar reward setups. Real-world experiments with an ANYmal-D quadruped robot confirm our policy's capability to generalize animal-like locomotion skills to complex terrains, demonstrating smooth and efficient locomotion and local navigation performance amidst challenging terrains with obstacles.

Figures

Figures reproduced from arXiv: 2505.16084 by the authors.

Figure 1
Figure 1. ANYmal-D hardware experiments in indoor/outdoor terrains (stairs, random blocks, high obstacles). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our training framework for quadruped locomotion and local navigation: (1) retarget [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Actual vs. reference joint positions for the low-level policy on flat ground. Dashed lines are reference [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Comparison of high-level joint residuals for the HFE joint on the left front (LF) and right front (RF) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Variance in joint residuals of LF-HFE joint, as observed in the experiment depicted in Fig. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Goal-reaching success rate on stairs and sloped terrains across different difficulty levels, with the left [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Average deviation of actual joint positions from reconstructed reference trajectories with FLD decoder [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Real world deployment of the high-level student policy. The sequence of images illustrates the ability [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Overview of the training environment and terrain configuration for high-level policy. The setup [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Footfall sequence of motion skills learned in low-level policy. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Overview of (a) pyramid stairs and (b) pyramid slope terrains with progressively increasing difficulty [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Energy efficiency comparison among three models: (i) the baseline RL model trained from scratch, [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why

    cs.LG 2025-07 conditional novelty 3.0 of 10

    Feature-based and GAN-based imitation learning should be selected by task priorities (fidelity, diversity, interpretability, adaptability), not by paradigm loyalty.

Reference graph

Works this paper leans on

36 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots. Science Robotics, 4(26):eaau5872, 2019. doi:10.1126/scirobotics.aau5872

  2. [2]

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

  3. [3]

    Zhuang, S

    Z. Zhuang, S. Yao, and H. Zhao. Humanoid parkour learning. In 8th Annual Conference on Robot Learning, 2024

  4. [4]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, and D. Pathak. Extreme parkour with legged robots. arXiv preprint arXiv:2309.14341, 2023

  5. [5]

    Zhang, N

    C. Zhang, N. Rudin, D. Hoeller, and M. Hutter. Learning agile locomotion on risky terrains, 2024

  6. [6]

    J. Ren, T. Huang, H. Wang, Z. Wang, Q. Ben, J. Pang, and P. Luo. Vb-com: Learning vision-blind composite humanoid locomotion against deficient perception, 2025

  7. [7]

    Z. Fu, T. Z. Zhao, and C. Finn. Mobile ALOHA: Learning bimanual mobile manipulation using low-cost whole-body teleoperation. In 8th Annual Conference on Robot Learning , 2024

  8. [8]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023

Show all 36 references
  1. [9]

    T. He, J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbabu, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. J. Fan, Y . Zhu, C. Liu, and G. Shi. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills.arXiv preprint ...

  2. [10]

    L. Han, Q. Zhu, J. Sheng, C. Zhang, T. Li, Y . Zhang, H. Zhang, Y . Liu, C. Zhou, R. Zhao, J. Li, Y . Zhang, R. Wang, W. Chi, X. Li, Y . Zhu, L. Xiang, X. Teng, and Z. Zhang. Lifelike agility and play in quadrupedal robots using reinforcement learning and generative pre-traine...

  3. [11]

    V ollenweider, M

    E. V ollenweider, M. Bjelonic, V . Klemm, N. Rudin, J. Lee, and M. Hutter. Advanced skills through multiple adversarial motion priors in reinforcement learning, 2022

  4. [12]

    C. Li, S. Blaes, P. Kolev, M. Vlastelica, J. Frey, and G. Martius. Versatile skill control via self-supervised adversarial imitation of unlabeled mixed motions. In 2023 IEEE international conference on robotics and automation (ICRA) , pages 2944–2950. IEEE, 2023

  5. [13]

    C. Li, M. Vlastelica, S. Blaes, J. Frey, F. Grimminger, and G. Martius. Learning agile skills via adversarial imitation of rough partial demonstrations. In Conference on Robot Learning, pages 342–352. PMLR, 2023

  6. [14]

    J. Wu, G. Xin, C. Qi, and Y . Xue. Learning robust and agile legged locomotion using adversarial motion priors. IEEE Robotics and Automation Letters , 8(8):4975–4982, 2023. doi:10.1109/ LRA.2023.3290509

  7. [15]

    Smith, J

    L. Smith, J. C. Kew, T. Li, L. Luu, X. B. Peng, S. Ha, J. Tan, and S. Levine. Learning and adapting agile locomotion skills by transferring experience, 2023

  8. [16]

    X. B. Peng, E. Coumans, T. Zhang, T.-W. E. Lee, J. Tan, and S. Levine. Learning agile robotic locomotion skills by imitating animals. In Robotics: Science and Systems , 07 2020. doi:10.15607/RSS.2020.XVI.064

  9. [17]

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Trans. Graph., 37(4):143:1– 143:14, July 2018. ISSN 0730-0301. doi:10.1145/3197517.3201311

  10. [18]

    C. Li, E. Stanger-Jones, S. Heim, and S. Kim. Fld: Fourier latent dynamics for structured motion representation and learning. arXiv preprint arXiv:2402.13820, 2024

  11. [19]

    Watanabe, C

    R. Watanabe, C. Li, and M. Hutter. Dfm: Deep fourier mimic for expressive dance motion learning, 2025

  12. [20]

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Trans. Graph. , 40(4), July 2021. doi:10.1145/3450626.3459670

  13. [21]

    Escontrela, X

    A. Escontrela, X. B. Peng, W. Yu, T. Zhang, A. Iscen, K. Goldberg, and P. Abbeel. Adversarial motion priors make good substitutes for complex reward functions. 2022 ieee. In International Conference on Intelligent Robots and Systems (IROS) , volume 2, 2022

  14. [22]

    X. B. Peng. Acquiring Motor Skills Through Motion Imitation and Reinforcement Learning . PhD thesis, EECS Department, University of California, Berkeley, Dec 2021

  15. [23]

    Z. Luo, J. Cao, J. Merel, A. Winkler, J. Huang, K. M. Kitani, and W. Xu. Universal humanoid motion representations for physics-based control. In The Twelfth International Conference on Learning Representations, 2024

  16. [24]

    X. B. Peng, Y . Guo, L. Halper, S. Levine, and S. Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), July 2022

  17. [25]

    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. doi:10.1126/scirobotics.adi7566. 10

  18. [26]

    Wellhausen and M

    L. Wellhausen and M. Hutter. Rough terrain navigation for legged robots using reachability planning and template learning. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6914–6921, 2021. doi:10.1109/IROS51168.2021.9636358

  19. [27]

    F. Yang, C. Wang, C. Cadena, and M. Hutter. iplanner: Imperative path planning, 2023

  20. [28]

    Zhang, J

    C. Zhang, J. Jin, J. Frey, N. Rudin, M. Mattamala, C. Cadena, and M. Hutter. Resilient legged local navigation: Learning to traverse with compromised perception end-to-end. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 34–41, 2024. doi:10.1109...

  21. [29]

    J. Lee, M. Bjelonic, A. Reske, L. Wellhausen, T. Miki, and M. Hutter. Learning robust autonomous navigation and locomotion for wheeled-legged robots. Science Robotics, 9(89): eadi9641, 2024. doi:10.1126/scirobotics.adi9641

  22. [30]

    Rudin, D

    N. Rudin, D. Hoeller, M. Bjelonic, and M. Hutter. Advanced skills by learning locomotion and local navigation end-to-end, 2022

  23. [31]

    Zhang, S

    H. Zhang, S. Starke, T. Komura, and J. Saito. Mode-adaptive neural networks for quadruped motion control. ACM Trans. Graph., 37(4), July 2018. ISSN 0730-0301. doi:10.1145/3197517. 3201366

  24. [32]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomotion over challenging terrain. Science Robotics, 5(47):eabc5986, 2020. doi:10.1126/scirobotics. abc5986

  25. [33]

    D. Chen, B. Zhou, V . Koltun, and P. Krähenbühl. Learning by cheating. In L. P. Kaelbling, D. Kragic, and K. Sugiura, editors, Proceedings of the Conference on Robot Learning , volume 100 of Proceedings of Machine Learning Research, pages 66–75. PMLR, 30 Oct–01 Nov 2020

  26. [34]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State. Isaac gym: High performance gpu-based physics simulation for robot learning, 2021

  27. [35]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms, 2017. Appendix 7 Training Details 7.1 Details for FLD Model Training As mentioned in Sec. 3.2, we train FLD model and use the FLD encoder to train low-level policy. The tra...

  28. [2022]

    doi:10.1126/scirobotics.abk2822

Pith tools

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