Pith. sign in

REVIEW 3 major objections 6 minor 25 references

UniLegs: Universal Multi-Legged Robot Control through Morphology-Agnostic Policy Distillation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single distilled Transformer student controls five trained robot morphologies at 94.47% of teacher performance and a never-seen unit at 72.64%, transferring to real hardware without fine-tuning.

desk verdict A plausible distillation framework with interesting numbers, but the missing mechanism for variable joint spaces makes the core claim unverifiable in its current form. read the letter →

arxiv 2507.22653 v2 pith:TDLG4PGG submitted 2025-07-30 cs.RO

Weijie Xi , Zhanxiang Cao , Chenlin Ming , Jianying Zheng , Guyue Zhou This is my paper · ORCID
classification cs.RO
keywords multi-leggedrobotcontrolpolicydistillationmorphology-agnosticTransformerteacher-studentframeworkzero-shotgeneralizationleggedlocomotionsim-to-realtransfer
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

This paper argues that a robot controller can be both specialized and universal if optimization and generalization are separated in time. The authors first train one reinforcement-learning teacher per robot morphology, then distill all teachers into a single Transformer-based student that takes a morphology descriptor as an additional input. Across five training morphologies—three quadrupeds plus six- and eight-legged robots—the distilled Transformer reaches 94.47% of each teacher's reward, and on a held-out Unitree Go2 quadruped it reaches 72.64% without any fine-tuning. The same distilled policy is then installed on a physical Go2 and walks stably. If this holds, legged robots do not need a bespoke controller for every new body.

What carries the argument

The load-bearing mechanism is a two-stage teacher–student distillation pipeline. In stage one, each morphology $M_i$ trains an independent PPO teacher with asymmetric actor–critic and privileged information such as base linear velocity and terrain height maps, so each teacher learns the optimal behavior for its own body. In stage two, a single student $\pi_\theta(a|o,m_i)$ imitates all teachers by minimizing the mean-squared error between its output and each teacher's action on a dataset of one million state–action pairs per morphology; the morphology descriptor $m_i$ is concatenated with the observation so one network can specialize by input. The student's Transformer uses three encoder layers with eight attention heads and a 256-dimensional embedding, matched in parameter count to an MLP baseline. Attention is the component that lets the same network dynamically reweight joint relationships for different bodies.

What would settle it

Train the pipeline with the morphology descriptor removed or randomly corrupted at test time and measure Go2 normalized reward; if performance barely changes, the descriptor is not carrying morphology identity and the universal-controller claim needs re-examination. Alternatively, hold out a morphology whose joint count lies outside the 12-to-24 range seen in training and test whether the policy still acts sensibly.

Watch

Extended reading notes

Core claim

The central discovery is that a universal multi-leg controller does not have to sacrifice per-robot optimality, because the trade-off lives in the training procedure rather than in the policy. By first letting each morphology converge to its own optimal teacher policy and only then distilling those policies into a shared student, the method preserves morphology-specific behaviors—94.47% of teacher reward averaged over training morphologies and 72.64% on an unseen Go2—whereas joint multi-morphology training and MLP students score lower. The Transformer student consistently beats the parameter-matched MLP baseline on every morphology, by 4.02 points on training morphologies and 3.35 points on Go2, which the authors attribute to self-attention modeling pairwise joint relationships that transfer across kinematic structures. The paper further claims the distilled Transformer policy transfers from simulation to a physical Go2 without fine-tuning, exhibiting stable gaits and command tracking.

Load-bearing premise

The approach depends on a single student network being able to read and output different numbers of joint values for different robots using one morphology descriptor, and the paper does not state how this variable-size input and output is encoded, padded, or masked.

Editorial extensions

If this is right

  • Adding a new robot to the fleet requires retraining only its teacher and re-running distillation, rather than re-optimizing a controller for every morphology.
  • A single onboard policy can serve deployments that mix quadrupeds, hexapods, and octopods, since the same weights work across different joint counts.
  • The 72.64% zero-shot figure sets the current bar for morphology-agnostic transfer in this pipeline, and the paper's own comparison shows Transformer attention consistently beats the matched MLP by several points on every body, training and unseen.
  • Because the student is trained to reproduce teacher actions, the 94.47% and 72.64% numbers also bound how much of the specialized behavior this distillation scheme can preserve.

Reading between the lines

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

  • Editorial inference: the paper does not specify how the student network handles variable observation and action dimensions across different joint counts; if that handled poorly, the universal controller claim could fail for morphologies with joint counts far outside the 12-to-24 range seen in training.
  • Editorial inference: since the student is trained by mean-squared-error imitation of fixed teacher policies, it cannot exceed teacher behavior; any further performance gains would require better teachers or a subsequent reinforcement-learning fine-tuning phase.
  • Editorial inference: a natural testable extension is to hold out a biped, a robot with asymmetric leg counts, or another morphology whose joint count lies far outside the training range, to probe whether the morphology descriptor generalizes or simply interpolates between trained body plans.
  • Editorial inference: the consistent Transformer advantage over the MLP student suggests that attention is doing real representational work for morphology transfer, not just adding capacity, since both architectures are parameter-matched; ablating the morphology descriptor would separate the contribution of the descriptor from the contribution of attention.
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 / 6 minor

Summary. The paper proposes UniLegs, a two-stage teacher-student framework for multi-morphology legged locomotion control. In the first stage, per-morphology teacher policies are trained with PPO using privileged information for five morphologies: Unitree A1, Aliengo, ANYmal, a six-legged robot, and an eight-legged robot. In the second stage, teacher action outputs are distilled offline, via a mean-squared-error loss, into a single student policy conditioned on a morphology descriptor m (Eqs. (5) and (8)); two student architectures are compared, a Transformer and an MLP intended to have equal parameter counts. Performance is reported as rewards normalized by each morphology's teacher policy (Table III), yielding 94.47% average teacher performance for the Transformer on training morphologies and 72.64% on an unseen Unitree Go2, versus 90.45% and 69.29% for the MLP. The paper also reports deployment of the distilled Transformer student on a physical Go2 without fine-tuning. The central claims are that distillation preserves morphology-specific optimal behavior, that the Transformer outperforms the MLP baseline, and that the framework zero-shot generalizes to unseen robot morphologies.

Significance. If the results are reproducible, the two-stage design is a sensible and potentially useful contribution: it decouples per-morphology optimization (which is sample-efficient and parallelizable) from cross-morphology generalization. The evaluation protocol is also honest in two respects: scores are normalized against per-morphology expert teachers rather than raw returns, and the Go2 expert is trained for evaluation only, so the 72.64% zero-shot figure is a genuine generalization measurement. I do not regard the reward-based evaluation as circular; evaluating a distilled student under the same task reward used to train the teachers is standard practice for distillation benchmarks. The paper makes no code or weights available, however, so reproducibility rests entirely on the textual specification, which is incomplete at exactly the point where the novelty lies (the variable-dimensionality mechanism). The headline effects are also small and unreplicated (e.g., 0.9843 vs 0.9814 on the eight-legged robot), so the Transformer-superiority claim cannot currently be assessed.

major comments (3)
  1. [III-C-2, Fig. 1, Eqs. (5) and (8)] The central mechanism of the claimed contribution — how one student network handles morphologies with different observation and action dimensionalities — is never specified. Figure 1 shows student inputs of dimension 45/63/81 for the 12/18/24-joint robots and outputs of dimension 12/18/24, and Section III-C-2 says only that the MLP maps 'concatenated morphology descriptors and state observations' to actions, while the Transformer is described by embedding width, head count, and three encoder layers. The paper never defines what the morphology descriptor m is, how variable-size observations are tokenized, padded, or masked into the shared 256-dimensional Transformer, how variable-size action vectors are emitted (e.g., per-morphology output heads), or how Eq. (8) aligns teacher and student action vectors of different lengths. The textual description and Figure 1 are even discrepant: the figure shows separate encoder/decoder paths per morphology, which the text does not describe. This gap is load-bearing for the zero-shot claim: if m is a categorical or learned identifier, the student never saw a Go2 descriptor during training and the zero-shot result would require an additional unexplained mechanism; if m is a kinematic descriptor, it must be defined and shown to be computable for Go2. As written, the forward pass is underdetermined and Table III cannot be reproduced from the paper alone. I regard this as the primary issue: the mechanism in question is the paper's stated contribution.
  2. [I-C (Contribution 1) and Table III] The headline comparative claim — that the framework 'significantly outperform[s] traditional multi-morphology training approaches' — is never tested. Table III compares only the per-morphology teachers (which define the normalization baseline) against the two distillation students; there is no baseline of a single policy trained jointly by RL across all five morphologies, which is precisely the approach the introduction argues against, and there is no comparison with any prior cross-morphology method (graph-based or Transformer-based). The claim in Contribution 2 of 'state-of-the-art performance' is likewise asserted without any comparison. Either add a joint multi-morphology RL training baseline (and ideally one prior cross-morphology method) to Table III, or remove these unsupported claims from the abstract, introduction, and conclusions.
  3. [Table III and IV-A-3] All headline numbers are single point estimates with no variance information. Table III reports one normalized reward per morphology per method, and Section IV-A-3 does not state the evaluation protocol: number of episodes, random seeds, command distributions, or terrain conditions used for the normalized reward scores, nor whether the scores come from closed-loop rollouts. Several of the claimed architectural differences are small (0.9843 vs 0.9814 for the eight-legged robot; 0.7264 vs 0.6929 on Go2), so without means and standard deviations over multiple teacher and student training seeds the paper's central claim that Transformer-based distillation 'consistently outperforms' the MLP cannot be distinguished from noise. Please report mean plus/minus standard deviation over at least 3-5 seeds for both distillation students and the evaluation rollout protocol.
minor comments (6)
  1. [IV-B-4 and Fig. 3] The real-world deployment is reported only qualitatively ('stable gait patterns', 'effective locomotion capabilities'); since hardware deployment is one of the three stated contributions, please add basic quantitative results (e.g., number of trials, command-tracking error or walking speed, failure modes) or a URL to an accessible video.
  2. [III-C-2] Parameter parity between the Transformer and the MLP is asserted but no parameter counts are given for either architecture; please report the actual numbers so this claim can be verified.
  3. [III-B-2 and Fig. 1] The paper does not state whether the student receives the same H=5 observation history as the teacher; Figure 1's '6x45' annotation suggests six stacked frames, but Section III-C does not describe the student's observation construction or its handling of history. Please clarify.
  4. [Table II] The foot-clearance term is typeset in a garbled way ('(pdes f,z,k - pf,z,k)2 * vf,xy,k'); please restore the equation, define all symbols, and check the remaining rows for the same typesetting problem.
  5. [III-B-2] 'FKBE leg configuration' is an unexplained acronym, and the number of feet n_f is introduced but never used; please define or remove both.
  6. [IV-B-3] The paper attributes the large morphology-dependent spread in transfer success (80.71% for A1 versus 99.36% for the six-legged robot) to morphology complexity, but offers no analysis of why A1 transfers so much worse than the other three quadrupeds; a brief discussion would strengthen the distillation-effectiveness argument.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the distillation results measure imitation fidelity against held-out teacher baselines, not a quantity forced by construction.

full rationale

All load-bearing quantities are externally grounded. Teacher policies are optimized by PPO (Eq. 4) with a fixed reward (Table II); this 'optimality' is an input assumption, not a consequence of the student equations. The student is trained by Eq. (8) to minimize MSE against teacher actions, and Table III reports the ratio of student reward to teacher reward. This is a standard distillation benchmark: the 94.47% and 72.64% figures are measurements of imitation quality, not outputs derived from the method's own definitions. The Go2 evaluation is not circular because the Go2 expert policy is 'used only for evaluation and is not included in the distillation process' (Sec. IV-A-3), so the zero-shot score is not a fitted parameter recycled as a prediction. Reference [19], which includes co-author G. Zhou, is cited only as general adaptive-learning background and is not load-bearing. The paper's main weakness is under-specification (how the Transformer encodes/pads variable joint dimensions and what the morphology descriptor m_i contains), which threatens reproducibility but is not an identity between an input and a claimed prediction. No self-definitional, fitted-input, or self-citation reduction is present.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three implicit assumptions: teacher optimality, the sufficiency of action-space MSE as a distillation objective, and the ability of a single architecture to handle variable joint counts. None of these are empirically scrutinized in the paper.

assumptions (3)
  • domain assumption Teacher policies trained with PPO are close to optimal for their own morphology.
    The entire evaluation normalizes student performance by teacher reward, so the 'optimal baseline' being 1.0 assumes the teachers are near-optimal; no comparison against other training methods is given. Section IV-B-1.
  • domain assumption MSE imitation of teacher actions is sufficient to recover the teacher's locomotion behavior and reward.
    The distillation loss (Eq. 8) minimizes action discrepancy; the paper does not analyze whether small action errors translate to large reward drops, though the Go2 drop to 72.64% suggests sensitivity. Section III-C-1.
  • domain assumption A single fixed-size Transformer can represent policies for different joint counts via a morphology descriptor.
    The paper never specifies how variable action/observation dimensions are handled; this assumption underlies the 'universal policy' claim. Section III-C-2 and Fig. 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniLegs: Universal Multi-Legged Robot Control through Morphology-Agnostic Policy Distillation." pith.science (2026). https://pith.science/paper/TDLG4PGG

@misc{pith2026250722653,
  author       = {Pith},
  title        = {Pith review of: UniLegs: Universal Multi-Legged Robot Control through Morphology-Agnostic Policy Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TDLG4PGG}},
  note         = {Machine review of arXiv:2507.22653}
}
read the original abstract

Developing controllers that generalize across diverse robot morphologies remains a significant challenge in legged locomotion. Traditional approaches either create specialized controllers for each morphology or compromise performance for generality. This paper introduces a two-stage teacher-student framework that bridges this gap through policy distillation. First, we train specialized teacher policies optimized for individual morphologies, capturing the unique optimal control strategies for each robot design. Then, we distill this specialized expertise into a single Transformer-based student policy capable of controlling robots with varying leg configurations. Our experiments across five distinct legged morphologies demonstrate that our approach preserves morphology-specific optimal behaviors, with the Transformer architecture achieving 94.47% of teacher performance on training morphologies and 72.64% on unseen robot designs. Comparative analysis reveals that Transformer-based architectures consistently outperform MLP baselines by leveraging attention mechanisms to effectively model joint relationships across different kinematic structures. We validate our approach through successful deployment on a physical quadruped robot, demonstrating the practical viability of our morphology-agnostic control framework. This work presents a scalable solution for developing universal legged robot controllers that maintain near-optimal performance while generalizing across diverse morphologies.

Figures

Figures reproduced from arXiv: 2507.22653 by the authors.

Figure 1
Figure 1. Morphology-Agnostic Control Framework. Our approach consists of two stages: (1) Morphology-specific teacher training where individual policies are learned for different robot morphologies (e.g., Aliengo, A1) using reinforcement learning with privileged information; (2) Cross-morphology knowledge distillation where a Transformer-based universal student policy learns to imitate all teacher policies using only base inf… view at source ↗
Figure 2
Figure 2. Training curves of teacher policies across different robot morphologies. The plot shows the mean reward as a function of training iterations for five distinct robot types: A1, Aliengo, Anymal_b, 6-legged, and 8-legged robots. All morphologies demonstrate stable convergence after 3000 iterations, with the multi-legged robots (6-legged and 8-legged) achieving higher final rewards compared to quadrupedal robots. 2) Per… view at source ↗
Figure 3
Figure 3. Real-world deployment of the distilled policy on a Unitree Go2 quadruped robot. The robot demonstrates stable locomotion capabilities in a laboratory environment, validating the sim-to-real transferability of our morphology-agnostic control approach. The physical experiments demonstrated notable consis￾tency with our simulation results, with the robot exhibiting stable gait patterns and effective locomotion capabili… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 13 canonical work pages

  1. [1]

    Learning robust perceptive locomotion for quadrupedal robots in the wild,

    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, vol. 7, no. 62, p. eabk2822, 2022

    work page 2022

  2. [2]

    Walk these ways: Tuning robot control for generalization with multiplicity of behavior,

    G. B. Margolis and P. Agrawal, “Walk these ways: Tuning robot control for generalization with multiplicity of behavior,” in Conference on Robot Learning . PMLR, 2023, pp. 22–31

    2023

  3. [3]

    Adapt- ing rapid motor adaptation for bipedal robots,

    A. Kumar, Z. Li, J. Zeng, D. Pathak, K. Sreenath, and J. Malik, “Adapt- ing rapid motor adaptation for bipedal robots,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022, pp. 1161–1168

    work page 2022

  4. [4]

    Metamorph: Learning universal controllers with transformers,

    A. Gupta, L. Fan, S. Ganguli, and L. Fei-Fei, “Metamorph: Learning universal controllers with transformers,” arXiv preprint arXiv:2203.11931, 2022

    arXiv 2022

  5. [5]

    Nervenet: Learning structured policy with graph neural networks,

    T. Wang, R. Liao, J. Ba, and S. Fidler, “Nervenet: Learning structured policy with graph neural networks,” in International conference on learning representations, 2018

    2018

  6. [6]

    My body is a cage: the role of morphology in graph-based incompatible control,

    V . Kurin, M. Igl, T. Rocktäschel, W. Boehmer, and S. Whiteson, “My body is a cage: the role of morphology in graph-based incompatible control,” arXiv preprint arXiv:2010.01856 , 2020

    arXiv 2010

  7. [7]

    Multi- embodiment legged robot control as a sequence modeling problem,

    C. Yu, W. Zhang, H. Lai, Z. Tian, L. Kneip, and J. Wang, “Multi- embodiment legged robot control as a sequence modeling problem,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 7250–7257

    work page 2023

  8. [8]

    Cross-domain policy adaptation via value-guided data filtering,

    K. Xu, C. Bai, X. Ma, D. Wang, B. Zhao, Z. Wang, X. Li, and W. Li, “Cross-domain policy adaptation via value-guided data filtering,” Ad- vances in Neural Information Processing Systems, vol. 36, pp. 73 395– 73 421, 2023

    work page 2023

Show all 25 references
  1. [9]

    Learning robot manipulation from cross-morphology demonstration,

    G. Salhotra, I. Liu, C. Arthur, and G. Sukhatme, “Learning robot manipulation from cross-morphology demonstration,” arXiv preprint arXiv:2304.03833, 2023

    2023 arXiv

  2. [10]

    One policy to control them all: Shared modular policies for agent-agnostic control,

    W. Huang, I. Mordatch, and D. Pathak, “One policy to control them all: Shared modular policies for agent-agnostic control,” in International Conference on Machine Learning . PMLR, 2020, pp. 4455–4464

    2020

  3. [11]

    Graph networks as learnable physics engines for inference and control,

    A. Sanchez-Gonzalez, N. Heess, J. T. Springenberg, J. Merel, M. Ried- miller, R. Hadsell, and P. Battaglia, “Graph networks as learnable physics engines for inference and control,” in International conference on machine learning . PMLR, 2018, pp. 4470–4479

    2018

  4. [12]

    Trakdis: A transformer-based knowledge distillation approach for visual reinforcement learning with application to cloth manipulation,

    W. Chen and N. Rojas, “Trakdis: A transformer-based knowledge distillation approach for visual reinforcement learning with application to cloth manipulation,” IEEE Robotics and Automation Letters , vol. 9, no. 3, pp. 2455–2462, 2024

    2024

  5. [13]

    Towards multi-morphology controllers with diversity and knowledge distillation,

    A. Mertan and N. Cheney, “Towards multi-morphology controllers with diversity and knowledge distillation,” in Proceedings of the Genetic and Evolutionary Computation Conference , 2024, pp. 367– 376

    2024

  6. [14]

    Hierarchically decoupled imitation for morphological transfer,

    D. Hejna, L. Pinto, and P. Abbeel, “Hierarchically decoupled imitation for morphological transfer,” in International Conference on Machine Learning. PMLR, 2020, pp. 4159–4171

    2020

  7. [15]

    Third-person visual imitation learning via decoupled hierarchical controller,

    P. Sharma, D. Pathak, and A. Gupta, “Third-person visual imitation learning via decoupled hierarchical controller,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  8. [16]

    Learning dexterous in-hand manipulation,

    O. M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. Mc- Grew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray et al., “Learning dexterous in-hand manipulation,” The International Journal of Robotics Research , vol. 39, no. 1, pp. 3–20, 2020

    2020

  9. [17]

    Learning to walk in minutes using massively parallel deep reinforcement learning,

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

    2022

  10. [18]

    Rma: Rapid motor adaptation for legged robots,

    A. Kumar, Z. Fu, D. Pathak, and J. Malik, “Rma: Rapid motor adaptation for legged robots,” arXiv preprint arXiv:2107.04034, 2021

    2021 arXiv

  11. [19]

    When to trust your simulator: Dynamics-aware hybrid offline-and-online reinforcement learning,

    H. Niu, Y . Qiu, M. Li, G. Zhou, J. Hu, X. Zhan et al. , “When to trust your simulator: Dynamics-aware hybrid offline-and-online reinforcement learning,” Advances in Neural Information Processing Systems, vol. 35, pp. 36 599–36 612, 2022

    2022

  12. [20]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  13. [21]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain et al. , “Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,” in 2024 IEEE International Conference on Robotics and Aut...

    2024

  14. [22]

    A generalist agent,

    S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-Maron, M. Gimenez, Y . Sulsky, J. Kay, J. T. Springenberg et al., “A generalist agent,” arXiv preprint arXiv:2205.06175 , 2022

    2022 arXiv

  15. [23]

    Diffuseloco: Real-time legged locomotion control with diffusion from offline datasets,

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

    2024 arXiv

  16. [24]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,” Science robotics, vol. 5, no. 47, p. eabc5986, 2020

    2020

  17. [25]

    Moral: Learning morphologically adaptive locomotion controller for quadrupedal robots on challenging terrains,

    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 Au- tomation Letters, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.