Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Reference Free Platform Adaptive Locomotion for Quadrupedal Robots using a Dynamics Conditioned Policy

T0 review · 4 major / 5 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 A single reinforcement-learning policy can drive quadrupeds from 12 kilograms to 50 kilograms with no per-robot retraining.

desk verdict Solid empirical study of reference-set diversity, but the zero-shot transfer claim is weaker than advertised because the test robot is one of the training reference models. read the letter →

arxiv 2505.16042 v2 pith:J7UQYHXS submitted 2025-05-21 cs.RO

David Rytz (1) , Suyoung Choi (2) , Wanming Yu (1) , Wolfgang Merkt (1) , Jemin Hwangbo (2) , Ioannis Havoutis (1) ((1) Dynamic Robot Systems , Oxford Robotics Institute , University of Oxford
show 3 more authors
(2) RaiLab Department of Mechanical Engineering KAIST)
This is my paper · ORCID
classification cs.RO
keywords quadrupedallocomotionreinforcementlearningzero-shottransferdynamicsconditioningproceduralrobotgenerationsim-to-realplatform-adaptivecontrolvelocitytracking
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 tries to establish that one reinforcement-learning policy can control quadrupeds of very different sizes and dynamics without per-robot retraining, by conditioning every action on a latent estimate of the robot's current dynamics. The policy is trained in simulation on procedurally generated variants of four reference quadrupeds and then transferred zero-shot to unseen simulated robots and to a real 50 kg ANYmal C. The authors report that the set of reference models chosen during training is itself a performance lever: including diverse morphologies reduces velocity command tracking error by up to 30% compared with training on a single reference model. They also find that a morphology-aware estimator of robot properties tracks velocity commands more accurately on ANYmal C hardware than their GRU-based temporal dynamics encoder, even though both approaches transfer successfully.

What carries the argument

The load-bearing mechanism is the dynamics-conditioned observation loop: the GRU-based DIM condenses recent base and joint state history into a latent vector of robot dynamics, and the CM policy receives that vector as part of its observation when outputting joint position targets. Around this loop sits procedural robot generation: each reference model yields 50 randomised variants drawn from the parameter ranges in Table I, admitted only if the robot can stand collision-free for 2 seconds in the simulator. Actuator command latency modelling and actuator networks trained on real ANYmal data are also injected during training. Together these parts make the policy platform-adaptive without motion priors: the latent embedding lets one set of weights respond appropriately across robots, and the generation ranges determine whether that embedding covers the target platform.

What would settle it

Train the same two architectures on a reference set that leaves out ANYmal C-like body masses and leg lengths, then deploy on ANYmal C and measure heading velocity RMSE. If the error stays near the full-reference values of 0.1652 for PAL and 0.0885 for MorAL instead of degrading toward the single-reference values of 0.2193 and 0.1339, the paper's claim that diverse reference modelling drives the improvement would be wrong; the expected contrast is given in the paper's own Table V.

Watch

Extended reading notes

Core claim

Platform Adaptive Locomotion (PAL) is a single reference-free locomotion policy consisting of a Dynamics Inference Module (DIM), a gated recurrent unit that reads recent base and joint history and emits a latent vector of the robot's temporally local dynamics, and a Control Module (CM) that maps proprioceptive state plus that latent vector to twelve desired joint positions tracked by PD controllers. Training happens on procedurally generated quadrupeds obtained by randomising the kinematic and dynamic properties of reference models spanning the A1, Aliengo, ANYmal B and ANYmal C. The central discovery is that conditioning on dynamics enables zero-shot transfer across quadrupeds from 12 kg to 50 kg, and that the diversity of reference models used during training is a decisive ingredient: the full reference set reduces hardware velocity tracking error by up to 30% relative to single-reference training, and the paper traces this to exposure to realistic mass distributions. On the ANYmal C hardware, the morphology-estimating MorAL architecture proved the more accurate controller, so the paper's contribution is evidence about which design choices matter rather than a claim that PAL beats every baseline.

Load-bearing premise

The load-bearing premise is that the hand-picked kinematic and dynamic randomisation ranges, together with the rule that a generated robot must stand collision-free for 2 seconds to be admitted, cover the real dynamics of every target robot, including the ANYmal C used in the hardware tests.

Editorial extensions

If this is right

  • A single trained policy can be deployed across quadrupeds from the 12 kg A1 to the 50 kg ANYmal C without per-robot retraining, including zero-shot transfer in simulation and on hardware.
  • Selecting multiple diverse reference quadrupeds for training reduces hardware velocity tracking error by up to 30% compared with a single-reference baseline, so reference-set choice is a controllable design parameter.
  • Morphology-aware conditioning yields lower velocity tracking error on ANYmal C than pure temporal dynamics encoding, indicating that architecture still matters once dynamics conditioning is in place.
  • Explicit modelling of actuation latency was required to move from simulation to the real ANYmal C, confirming that actuator response is one of the main sim-to-real gaps.

Reading between the lines

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

  • If the trend holds, future universal legged controllers could be specified by a parameter distribution rather than a robot identity, so a new platform would be added by checking that its dynamics fall inside the trained range.
  • The two conditioning strategies appear complementary, since the velocity-estimator results are similar across architectures; combining the GRU dynamics encoder with a morphology-property estimator is a natural next test.
  • The hand-picked randomisation ranges are themselves the real hypothesis, so a formal coverage test over the target robot's parameter space would predict when zero-shot transfer succeeds and when it fails.
  • The 30% improvement from reference diversity suggests that scaling a single reference model, as earlier work did, under-samples the mass and inertia regimes that matter for heavier robots.
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 Platform Adaptive Locomotion (PAL), a single reinforcement-learned controller for quadrupedal robots with diverse morphologies and dynamics. The policy is trained on procedurally randomized variants of several reference quadruped models (A1, Aliengo, ANYmal B, ANYmal C) and is conditioned on either a GRU-based latent dynamics embedding (DIM) or an explicit morphology estimator (MorAL-style). The authors compare PAL against MorAL and GenLoco baselines under different sets of training reference models, measuring simulation success rates under perturbations and hardware velocity-tracking performance on ANYmal C. The reported results show that training on a broader set of reference models improves robustness and tracking, that PAL achieves competitive simulation robustness on the smaller A1, and that the MorAL baseline achieves lower tracking error on ANYmal C hardware. The paper also documents a negative result: PAL does not outperform MorAL on the ANYmal C hardware task.

Significance. If the central claims were fully substantiated, the paper would make a useful contribution: a reference-free policy that transfers across a 12–50 kg range of quadrupeds without per-robot retraining, with explicit ablations over the choice of training reference models. The paper is commendably transparent about its negative hardware result for PAL versus MorAL, and it provides reproducible-looking training details and multi-seed learning curves in Figure 3. However, the significance is currently limited by the fact that the 'zero-shot transfer' claim is not actually demonstrated: the two robots evaluated in detail (A1 and ANYmal C) are themselves training reference models whose nominal parameters lie inside the sampled distributions, and the hardware comparison lacks trial counts, error bars, and ground-truth velocity references.

major comments (4)
  1. [IV.A, IV.B, Table I] The claim of zero-shot transfer to ANYmal C and A1 is not supported by the evidence, because both robots are used as reference models for generating the training distribution. Table I samples the base mass of ID 5 (ANYmal C) from U(18,50) kg, so the nominal 50 kg ANYmal C lies at the upper edge of the trained mass range, and ID 1 (A1) has its nominal mass inside U(2,28) kg. Section IV.A states that 'the specific quadrupeds were never included in any of the training sets, but instead used as a base model in which we change the sampled kinematic and dynamic parameters'; this is true only in the sense that the exact nominal parameter vector is not sampled, but the nominal configuration is still within or near the support of the training distribution. A leave-one-reference-out experiment (e.g., train on IDs {1,2,4} and test on nominal ANYmal C, or evaluate on ID 3, which is never used in the multi-ID training sets) is needed to substantiate the word 'zero-shot'. Without such an experiment, the hardware and simulation results are better described as interpolation over training reference models.
  2. [III-B.4, IV.C] The hardware transfer to ANYmal C is not zero-shot with respect to actuation dynamics. Section III-B.4 states that the actuator networks which randomly replace Equation (3) during training were trained on data collected on ANYmal B and C hardware. Since the hardware deployment target is ANYmal C, the policy's training distribution already includes the target robot's actuator characteristics. This should be acknowledged as a limitation, or the actuator networks should be retrained without ANYmal C data to preserve the zero-shot claim.
  3. [IV.D, Tables V and VI] The hardware RMSE tables contain no trial counts, per-run variance, or significance tests. The reported differences (e.g., PAL RMSE-X 0.1652 vs 0.2019, or MorAL RMSE-X 0.0885 vs 0.1339) may be within run-to-run variability on hardware, especially given the small number of conditions. Additionally, Table V compares the target velocity command with the learned velocity estimator, not with ground-truth base velocity, so estimator errors are conflated with tracking errors. The paper should report the number of trials, error bars, and ideally an independent motion-capture or onboard-state-estimator cross-check for at least one condition before claiming percentage improvements.
  4. [IV.A, III.E] The MorAL baseline is modified relative to the original method: the height map is removed and the action history length differs from the original MorAL policy. The paper does not provide an ablation quantifying the effect of these changes on MorAL's performance. Since the central hardware conclusion is that MorAL outperforms PAL, it should be shown that these modifications do not bias the comparison in either direction.
minor comments (5)
  1. [Abstract, V] The claim of 'reducing the velocity tracking error by up to 30%' is ambiguous about which controller and which baseline are being compared. In Table V, the 30+% reduction appears for MorAL (RMSE-X 0.0885 vs 0.1339), while PAL shows about 25% reduction (0.1652 vs 0.2193); the wording should be clarified.
  2. [IV.B] The text says 'The same early termination criteria defined in Section III-B are applied,' but Section III-B describes the robot admission filter, not early termination criteria; the early termination penalty is described later in Section III-G. Please correct the cross-reference.
  3. [Table III] The air-time reward expression is malformed: 'rair,i = −3 · ( if ||cdes||2 = 0 : −Ti,swing else: Ti,swing − 0.5' is missing a closing parenthesis and the intended sign convention is unclear.
  4. [Figure 4] The black 'Min/Max' training ranges for the horizontal force perturbation are not defined in Table I or in the text; please provide the corresponding sampling ranges for each perturbation type.
  5. [III.E] The sentence 'Compared to the MorAL policy of Luo et al. [14], we only use the state history of the past two control steps instead of the past five' is ambiguous: it is unclear whether this change applies to the PAL policy only or to both PAL and the implemented MorAL baseline. Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the policy is evaluated against external command-tracking benchmarks; self-citations are implementation details, and the ANYmal C training-reference issue is a generalization-validity concern, not a derivation loop.

full rationale

The paper's central claim—that a single dynamics-conditioned policy reduces velocity tracking error by up to 30% versus a baseline—is supported by hardware RMSE measurements (Table V) comparing PAL and MorAL against commanded velocities; these outputs are not constructed from the inputs. The policy is trained with domain randomization and then evaluated on command tracking; no fitted parameter is later renamed as a prediction. The main validity caveat is that ANYmal C (ID 5) is one of the reference base models used to sample the training distribution (Section III-B.1, Table I), so the hardware deployment is interpolation over the training support rather than a truly unseen-robot test. However, this is a limitation of the generalization claim, not circularity, because the test performance is not used to define or fit the policy or the conditioning module. The self-citations to prior work ([3], [25], [27]) provide actuator-network models, reward terms, and hyperparameter starting points; they are load-bearing only as engineering choices, and the head-to-head comparisons use externally defined baselines (GenLoco [15], MorAL [14]). No uniqueness theorem or ansatz is imported from the authors' prior work to forbid alternatives. Therefore no step reduces to its own input by construction.

Assumptions & free parameters 9 free parameters · 5 assumptions · 1 invented entities

The central generalization claims rest on a large set of hand-chosen reward weights, sampling ranges, and architecture sizes from Tables I, III, and IV, plus simulation and actuator models pulled from prior work. These are inputs, not derived quantities. The latent dynamics vector is an unvalidated internal representation, and the delay distribution is not fully specified.

free parameters (9)
  • Reward coefficients (13 terms, Table III) = rv=3, r_w=1.75, rR=-5, rh=-20, rb=-0.5, rq=-0.2, r_qdot=-3e-4, r_qddot=-2e-7, r_tau=-3.5e-5, rs1=-0.1, rs2=-0.05…
    Hand-tuned based on the authors' prior work [25]; they shape the learned gait and are not derived from first principles.
  • Action noise standard deviation sigma_a = 0.6
    Set by hand; controls exploration of the Gaussian policy.
  • GRU hidden state size h = 36
    Chosen as the dimension of the latent dynamics embedding; no ablation shown.
  • Observation history length = 2 control steps
    Reduced from 5 in MorAL; affects what the dynamics encoder and policy can infer.
  • Kinematic/dynamic randomization ranges (Table I) = Ranges for link offsets, masses, PD gains, torque limits, friction
    Hand-chosen intervals per reference robot; define the training distribution that the generalization claim depends on.
  • Command sampling ranges = v_x in [-1,1] m/s, v_y in [-0.75,0.75] m/s, omega_z in [-1.5,1.5] rad/s, duration 3-6 s
    Defines the task distribution; not justified by a task specification.
  • PPO hyperparameters (Table IV) = gamma=0.9962, batch 63000, minibatch 8, epochs 4, nenv=450, BPTT batch 50
    Hand-tuned based on [25]; not swept in this paper.
  • Robot admission filter = Must stand collision-free for 2 s
    Ad hoc feasibility filter; may exclude or include robots in ways that affect coverage.
  • Actuation command tracking delay distribution = Not quantified in text
    Delays are added and actuator networks [3] replace PD output, but the delay distribution parameters are not reported, making the sim-to-real setup hard to replicate.
assumptions (5)
  • domain assumption RaiSim rigid-body simulation with its contact model adequately approximates real quadruped dynamics for policy transfer.
    Used throughout Sec III-B/IV for all training and robustness claims; no independent validation against real robot dynamics beyond ANYmal C behavior.
  • domain assumption Actuator networks trained on ANYmal B and C hardware data capture the non-linear actuation dynamics of the target robots.
    Sec III-B.4; sim-to-real transfer hinges on this model.
  • domain assumption The learned base linear velocity estimator is accurate enough to serve as both policy input and evaluation metric.
    Sec III-E and Sec IV-E; hardware RMSE values assume the onboard comparison is a valid ground truth and the estimator is not gaming the command.
  • domain assumption A PD controller with the sampled Kp/Kd ranges can track the policy's joint targets sufficiently well.
    Eq (3) and Table I; if actuator saturations or delays break this assumption, the policy actions are not realized.
  • standard math PPO with the given hyperparameters converges to a policy representative of the reward design.
    Standard RL assumption; no proof of optimality.
invented entities (1)
  • Latent dynamics vector s_d (GRU hidden state, h=36)
    purpose: Condition the control policy on temporally local system dynamics so one policy can handle multiple robot morphologies.
    No independent measurement or prediction validates the latent content; its usefulness is only inferred from downstream task performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reference Free Platform Adaptive Locomotion for Quadrupedal Robots using a Dynamics Conditioned Policy." pith.science (2026). https://pith.science/paper/J7UQYHXS

@misc{pith2026250516042,
  author       = {Pith},
  title        = {Pith review of: Reference Free Platform Adaptive Locomotion for Quadrupedal Robots using a Dynamics Conditioned Policy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7UQYHXS}},
  note         = {Machine review of arXiv:2505.16042}
}
read the original abstract

This article presents Platform Adaptive Locomotion (PAL), a unified control method for quadrupedal robots with different morphologies and dynamics. We leverage deep reinforcement learning to train a single locomotion policy on procedurally generated robots. The policy maps proprioceptive robot state information and base velocity commands into desired joint actuation targets, which are conditioned using a latent embedding of the temporally local system dynamics. We explore two conditioning strategies - one using a GRU-based dynamics encoder and another using a morphology-based property estimator - and show that morphology-aware conditioning outperforms temporal dynamics encoding regarding velocity task tracking for our hardware test on ANYmal C. Our results demonstrate that both approaches achieve robust zero-shot transfer across multiple unseen simulated quadrupeds. Furthermore, we demonstrate the need for careful robot reference modelling during training: exposing the policy to a diverse set of robot morphologies and dynamics leads to improved generalization, reducing the velocity tracking error by up to 30% compared to the baseline method. Despite PAL not surpassing the best-performing reference-free controller in all cases, our analysis uncovers critical design choices and informs improvements to the state of the art.

Figures

Figures reproduced from arXiv: 2505.16042 by the authors.

Figure 1
Figure 1. Overall pipeline. A Dynamic Inference Module (DIM) in the form of a gated recurrent unit generates an approximate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the kinematic robot description utilized [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The average learning performance of three controller [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Success rate observed for various motion control policies for different perturbations and dynamics parameters for [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: PAL locomotion policy deployed on real ANYmal [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID 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. Sampling Strategies for Robust Universal Quadrupedal Locomotion Policies

    cs.RO 2025-10 conditional novelty 6.0 of 10

    A particle-filter-based adaptive sampling of morphologies and wide PD-gain randomization yields a single quadruped locomotion policy that transfers zero-shot to ANYmal hardware.

Reference graph

Works this paper leans on

30 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Dynamic Locomotion Through Online Nonlinear Motion Optimization for Quadrupedal Robots,

    C. D. Bellicoso, F. Jenelten, C. Gehring, and M. Hutter, “Dynamic Locomotion Through Online Nonlinear Motion Optimization for Quadrupedal Robots,” IEEE Robotics and Automation Letters , vol. 3, no. 3, pp. 2261–2268, 2018

    work page 2018

  2. [2]

    Crocoddyl: An Efficient and Versatile Framework for Multi-Contact Optimal Control,

    C. Mastalli, R. Budhiraja, W. Merkt, G. Saurel, B. Hammoud, M. Naveau, J. Carpentier, L. Righetti, S. Vijayakumar, and N. Mansard, “Crocoddyl: An Efficient and Versatile Framework for Multi-Contact Optimal Control,” in IEEE International Conference on Robotics and Automation , 2020, pp. 2536–2542

    work page 2020

  3. [3]

    Learning agile and dynamic motor skills for legged robots,

    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 , vol. 4, no. 26, 2019. [Online]. Available: https://www.science.org/doi/10.1126/scirobotics.aau5872

    doi:10.1126/scirobotics.aau5872 2019

  4. [4]

    RLOC: Terrain-Aware Legged Locomotion using Reinforcement Learning and Optimal Control

    S. Gangapurwala, M. Geisert, R. Orsolino, M. Fallon, and I. Havoutis, “RLOC: Terrain-Aware Legged Locomotion using Reinforcement Learning and Optimal Control,” IEEE Transactions on Robotics, vol. 38, no. 5, pp. 2908–2927, 2022. [Online]. Available: http://arxiv.org/abs/2012.03094

    work page Pith review arXiv 2022

  5. [5]

    Learning agile robotic locomotion skills by imitating animals,

    X. B. Peng, E. Coumans, T. Zhang, T. W. E. Lee, J. Tan, and S. Levine, “Learning agile robotic locomotion skills by imitating animals,” arXiv preprint arXiv:2004.00784, 2020

    arXiv 2004

  6. [7]

    Perceptive Locomotion in Rough Terrain – Online Foothold Optimization,

    F. Jenelten, T. Miki, A. E. Vijayan, M. Bjelonic, and M. Hutter, “Perceptive Locomotion in Rough Terrain – Online Foothold Optimization,” IEEE Robotics and Automation Letters , 2020. [Online]. Available: https://doi.org/10.1109/LRA.2020.3007427

    arXiv 2020

  7. [8]

    [Online]

    Unitree Robotics, “A1,” 2025. [Online]. Available: https://www. unitree.com/a1

    work page 2025

  8. [9]

    ANYbotics Introduces Sleek New ANYmal C Quadruped,

    E. Ackerman, “ANYbotics Introduces Sleek New ANYmal C Quadruped,” 2019. [Online]. Available: https://spectrum.ieee.org/ anybotics-introduces-sleek-new-anymal-c-quadruped

    work page 2019

Show all 30 references
  1. [10]

    Contact Invariant Model Learn- ing for Legged Robot Locomotion,

    R. Grandia, D. Pardo, and J. Buchli, “Contact Invariant Model Learn- ing for Legged Robot Locomotion,” IEEE Robotics and Automation Letters, vol. 3, no. 3, pp. 2291–2298, 2018

    2018

  2. [11]

    PILCO: A Model-Based and Data-Efficient Approach to Policy Search,

    M. P. Deisenroth and C. E. Rasmussen, “PILCO: A Model-Based and Data-Efficient Approach to Policy Search,” in Proceedings of the 28th International Conference on Machine Learning , 2011, pp. 465–472

    2011

  3. [12]

    World Models,

    D. Ha and J. Schmidhuber, “World Models,” arXiv preprint arXiv:1803.10122, 2018. [Online]. Available: http://arxiv.org/abs/ 1803.10122

    2018 arXiv

  4. [13]

    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

  5. [14]

    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 Automation Letters , pp. 1–8, 2024. [Online]. Available: https://ieeexplore.ieee.or...

    2024

  6. [15]

    GenLoco: Generalized Locomotion Controllers for Quadrupedal Robots,

    G. Feng, H. Zhang, Z. Li, X. B. Peng, B. Basireddy, L. Yue, Z. Song, L. Yang, Y . Liu, K. Sreenath, and S. Levine, “GenLoco: Generalized Locomotion Controllers for Quadrupedal Robots,” 2022. [Online]. Available: http://arxiv.org/abs/2209.05309

    2022 arXiv

  7. [16]

    ManyQuadrupeds: Learning a Single Locomotion Policy for Diverse Quadruped Robots,

    M. Shafiee, G. Bellegarda, and A. Ijspeert, “ManyQuadrupeds: Learning a Single Locomotion Policy for Diverse Quadruped Robots,” 2023. [Online]. Available: http://arxiv.org/abs/2310.10486

    2023 arXiv

  8. [17]

    MetaLoco: Universal Quadrupedal Locomotion with Meta-Reinforcement Learning and Motion Imitation,

    F. Zargarbashi, F. D. Giuro, J. Cheng, D. Kang, B. Sukhija, and S. Coros, “MetaLoco: Universal Quadrupedal Locomotion with Meta-Reinforcement Learning and Motion Imitation,” 2024. [Online]. Available: http://arxiv.org/abs/2407.17502

    2024 arXiv

  9. [18]

    One Policy to Run Them All: An End-to-end Learning Approach to Multi-Embodiment Locomotion,

    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,”

  10. [19]

    Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion,

    G. Ji, J. Mun, H. Kim, and J. Hwangbo, “Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 4630–4637, 2022

    2022

  11. [20]

    R. S. Sutton and A. G. Barto, Reinforcement Learning - An Introduc- tion. MIT Press, 1998

    1998

  12. [21]

    ANYmal - a highly mobile and dynamic quadrupedal robot,

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, and M. Hoepflinger, “ANYmal - a highly mobile and dynamic quadrupedal robot,” in IEEE/RSJ International Conference on I...

    2016

  13. [22]

    RaiSim – v1.1.7 documentation,

    RaiSim Tech, “RaiSim – v1.1.7 documentation,” 2025. [Online]. Available: https://raisim.com/

    2025

  14. [23]

    On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models,

    J. Schmidhuber, “On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models,” arXiv preprint arXiv:1511.09249 , 2015. [Online]. Available: http://arxiv.org/abs/1511.09249

    2015 arXiv

  15. [24]

    Deepim: Deep iterative matching for 6d pose estimation,

    Y . Li, G. Wang, X. Ji, Y . Xiang, and D. Fox, “Deepim: Deep iterative matching for 6d pose estimation,” in Proceedings of the European Conference on Computer Vision , 2018, pp. 683–698

    2018

  16. [25]

    Learning Low-Frequency Motion Control for Robust and Dynamic Robot Locomotion,

    S. Gangapurwala, L. Campanaro, and I. Havoutis, “Learning Low-Frequency Motion Control for Robust and Dynamic Robot Locomotion,” in IEEE International Conference on Robotics and Automation. arXiv, 2023, pp. 5085–5091. [Online]. Available: http://arxiv.org/abs/2209.14887

    2023 arXiv

  17. [26]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv, pp. 1–12, 2017

    2017

  18. [27]

    Guided Constrained Policy Optimization for Dynamic Quadrupedal Robot Locomotion,

    S. Gangapurwala, A. Mitchell, and I. Havoutis, “Guided Constrained Policy Optimization for Dynamic Quadrupedal Robot Locomotion,” IEEE Robotics and Automation Letters , vol. 5, no. 2, pp. 3642–3649, 2020

    2020

  19. [28]

    A focused backpropagation algorithm for temporal pattern recognition,

    M. C. Mozer, “A focused backpropagation algorithm for temporal pattern recognition,” in Backpropagation. Psychology Press, 2013, vol. 137

    2013

  20. [29]

    Learning Memory-Based Control for Human-Scale Bipedal Locomotion,

    J. Siekmann, S. Valluri, J. Dao, L. Bermillo, H. Duan, A. Fern, and J. Hurst, “Learning Memory-Based Control for Human-Scale Bipedal Locomotion,” 2020. [Online]. Available: http://arxiv.org/abs/ 2006.02402

    2020 arXiv

  21. [2023]

    Available: http://arxiv.org/abs/2309.05665

    [Online]. Available: http://arxiv.org/abs/2309.05665

    arXiv

  22. [2025]

    Available: http://arxiv.org/abs/2409.06366

    [Online]. Available: http://arxiv.org/abs/2409.06366

Pith tools

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