Pith. sign in

REVIEW 3 major objections 3 minor 84 references

A single onboard neural-network policy can run, jump, climb, and land on unstructured terrain at up to 6 meters per second, choosing gaits and skills on its own from camera and LiDAR inputs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A single onboard policy trained with 2D trajectory-optimization priors, transformer latent actions, and reinforcement learning enables a quadruped to autonomously select gaits and traverse unstructured terrain at up to 6 m/s.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A real architecture advance with convincing ablations; treat the headline real-world speeds as demo footage, not measured claims. the 3 major comments →

arxiv 2607.13579 v1 pith:WOKFTDVY submitted 2026-07-15 cs.RO cs.AIcs.LG

Agile perceptive multi-skill locomotion for quadrupedal robots in the wild

Jun-Gill Kang , Jaehyun Park , Tae-Gyu Song , Joon-Ha Kim , Seungwoo Hong , Hae-Won Park This is my paper
classification cs.RO cs.AIcs.LG
keywords quadruped locomotionreinforcement learningtrajectory optimizationaction pretrainingperceptive locomotiongait selectionsim-to-real transfertransformer VAE
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

APT-RL claims that a quadruped robot can traverse unstructured outdoor terrain — stairs, hurdles, stepping stones, gaps, high steps, logs — at speeds up to 4.25 m/s over a 60 cm step and 6 m/s during a stair descent, using a single onboard policy and only camera, LiDAR, and onboard compute. The method generates 15.5 hours of flat-ground sagittal-plane running motion in about 8 minutes via trajectory optimization, then learns a compact latent space and two gait-specific torque decoders with a transformer variational autoencoder. Reinforcement learning reuses those frozen decoders while jointly learning a small auxiliary action, and a perception student is distilled from privileged height maps. If the claims hold, high-speed multi-skill locomotion no longer requires motion-capture databases, online trajectory optimization, or manually designed high-level skill switching.

Core claim

One onboard policy autonomously selects trot or bound and switches skills to traverse real-world stairs, hurdles, stepping stones, gaps, logs, and high steps, with zero-shot sim-to-real transfer. Peak instantaneous speeds reach 4.25 m/s over a 60 cm step and 6 m/s descending a three-step stair, using only depth, LiDAR, and onboard compute. The enabler is a 2D trajectory-optimization dataset (15.5 hours of motion generated in 8 minutes) that trains a transformer VAE and two frozen gait-specific torque decoders; RL then adds a 12-dimensional auxiliary action and 2 Hz gait selection. Without the decoder torque, success collapses from 94.6% to 2.9% in simulation, indicating the prior is non-redu

What carries the argument

The APT module is the load-bearing component: a transformer-based variational autoencoder trained on 180,000 2D sagittal-plane trajectories generated by trajectory optimization with single rigid body dynamics and polynomial ground-reaction-force profiles. It produces a 16-dimensional latent state code and two gait-specific torque decoders (trot, bound), which stay frozen during RL. The policy outputs a 29-dimensional action: a latent action decoded by the chosen decoder, a 12-dimensional auxiliary action interpreted as a PD target, and a gait-selection logit switched at 2 Hz. The final joint torque is the decoder torque plus the auxiliary refinement, so the pretrained decoder supplies feedfo

Load-bearing premise

The load-bearing premise is that a torque decoder pretrained on flat-ground 2D sagittal-plane motions still provides an indispensable, non-redundant action prior for 3D obstacle traversal; if that flat data cannot cover the dynamics needed on uneven terrain, the auxiliary action would have to do all the work, which the paper's own ablation says is not possible.

What would settle it

On the real robot, on the same high-step course used in simulation, rerun the deployment with the pretrained decoder torque set to zero while keeping the auxiliary PD action active. The paper's simulation ablation predicts success dropping from 94.6% to 2.9%; if the robot instead clears the obstacles more than half the time, the claim that the 2D trajectory-optimization prior is non-redundant is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, high-speed multi-skill locomotion can be obtained without animal motion-capture data or online trajectory optimization during learning.
  • The pretrained torque decoder is the dominant control signal; removing it collapses simulated success on high-step terrain from 94.6% to 2.9%.
  • The same 2D flat-ground dataset supports multiple gaits and smooth autonomous switching, so skill diversity can come from cheap data rather than hand-tuned transition logic.
  • Depth plus long-range LiDAR, distilled from a privileged height-map teacher, is sufficient for zero-shot real-world perception; neither modality alone matches both.
  • The architecture is robot-agnostic and can be repurposed for other embodiments, per the paper's preliminary demonstrations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the pretrained torque prior is 2D sagittal-plane data, the same pipeline will likely need 3D or lateral motion data before fast turning and side-stepping reach the same agility; the paper states these as future directions.
  • The peak-speed numbers are instantaneous, measured around jumps and drop-downs, not sustained cruising speeds; they index dynamic capability rather than average traversal speed.
  • A practical co-design constraint surfaced in the paper: the LiDAR failed under impacts exceeding 10 g, so the vibration absorber was not optional; further speed gains may demand sensor-hardware changes as much as algorithmic ones.
  • The equivalence the paper shows between direct feedforward torque and an equivalent PD target suggests the contribution is the learned torque prior itself, not the control representation, so other robot controllers could adopt the same prior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper introduces APT-RL, a three-phase pipeline for quadruped locomotion: (1) generate large-scale 2D flat-ground motion datasets with trajectory optimization using a single rigid-body model and Bézier ground-reaction forces, (2) learn a Transformer-VAE latent space and gait-specific torque decoders, and (3) train an RL policy that outputs a latent action decoded to feedforward torque plus an auxiliary PD action, with a trainable gait-selection logit, followed by distillation from privileged height maps to onboard depth+LiDAR perception. The authors claim that the resulting single onboard policy autonomously selects gaits and motor skills, achieves zero-shot sim-to-real transfer, and enables high-speed traversal of stairs, steps, gaps, logs, and rough terrain, with reported instantaneous peak speeds of 4.25 m/s and 6 m/s on a stair drop. The central technical thesis is that 2D TO data provide a non-redundant torque prior for 3D agile locomotion.

Significance. If the real-world claims hold, this is a substantial advance: it would show that a flat-ground 2D trajectory-optimization prior, a shared latent action space, and a jointly learned auxiliary action can replace hierarchical skill selection and external motion capture, while enabling perceptive high-speed multi-skill locomotion in unstructured environments. The simulation evidence is genuinely strong in several respects: the ablations are well-structured (three seeds, 300 agents, AMP/HRL baselines, sensor ablation), and the feedforward-torque ablation (Fig. S8) directly supports the central mechanism, with success collapsing from 94.6% to 2.9% when the pretrained torque is removed. The data generation is unusually fast (180,000 trajectories in about 8 minutes), and the authors deposit data and figure-generation code in Zenodo, which supports reproducibility. The main gap is the real-world evidence: the headline speeds and robustness claims rest on single successful trials without attempt counts, failure rates, or uncertainty quantification, and the sim-to-real perception pipeline contains unquantified preprocessing steps.

major comments (3)
  1. [Materials and Methods: 'Distillation of exteroception latent using teacher-student framework'; Supplementary Hardware im] The headline quantitative claims (4.25 m/s and 6 m/s instantaneous peak speeds, and robust traversal of diverse obstacles) rest on single successful trials. The Statistical Analysis section explicitly states that no hypothesis tests, p-values, or confidence intervals were computed, and Fig. S2 shows a single camera-tracked trial for the 4.25 m/s claim. The 6 m/s value is derived by low-pass filtering body velocity relative to contact feet at 60 Hz, a method that can overestimate inertial speed during impact or slippery landings. No attempt counts or failure counts are reported for the urban, wild, or indoor deployments. This is load-bearing because the paper's central claim is 'robust traversal' and 'instantaneous peak speeds' in the real world. Please report the number of trials, success/failure counts, and a validated speed measurement (e.g., motion capture, GNSS, or an independent ine
  2. [Results: 'Fast perceptive quadruped locomotion...'; Fig. 1; Fig. 3] The zero-shot sim-to-real claim is qualified by an unquantified domain gap. The paper states that native LiDAR raycasting was unavailable in the simulator, so training used a 2.5D heightmap, and deployment required custom preprocessing to convert real LiDAR data into the simulated format. In addition, the custom mechanical vibration absorber is described as 'critical' for LiDAR stability during high-speed running, but its effect on measurement quality or downstream policy performance is not evaluated. This matters because the claim is that the policy operates 'using only onboard LiDAR, a depth camera, and onboard computation' in zero-shot fashion. Without quantifying how faithful the LiDAR preprocessing is and how sensor stability varies with the absorber, the perception module's contribution to the real-world robustness claim is unestablished. Please add a quantitative evaluation of the
  3. [Discussion; Fig. S8; Fig. S9] The narrative repeatedly states that the robot 'successfully traversed' or 'completed' courses and 'robustly' traversed diverse obstacles, but no success rates, failure counts, or repeat-run data are given for the real-world courses. For a paper whose central claim is 'robust traversal of diverse obstacles,' the absence of any repeated-run statistics in the physical world is a significant gap. The simulation ablations (Figs. 6C, 7, 8) are informative, but they do not by themselves establish real-world robustness. I ask the authors to provide a table of real-world runs per terrain type, including failures and near-failures, or to soften the language to 'demonstrated on single successful trials.'
minor comments (3)
  1. [Materials and Methods: 'Reinforcement learning...'] Minor typo: 'enabling smooth, high-speed multi-skill locomotion in deployed policy' should be 'in the deployed policy' or 'in deployment.'
  2. [Supplementary: 'Ablation study: Feedforward torque contribution'] The control law τ_input = τ_dec + PD(auxiliary) is later shown in Fig. S9 to be algebraically equivalent to a PD controller with a shifted reference. This is fine, but the main text describes the pretrained torque as the dominant 'non-redundant' component; the equivalence means the claim should be phrased as 'the pretrained signal is non-redundant as a reference/feedforward term,' not necessarily as a raw torque. Please clarify this nuance in the main text.
  3. [Supplementary: 'Ablation study: Dataset size...'] The ablation labels are not all self-explanatory. For example, 'Ours w/o latent RL policy' uses a pretrained encoder to map observations to latent actions, which is not an RL policy; please clarify the exact replacement in the caption or main text.

Circularity Check

0 steps flagged

No meaningful circularity: central empirical claims are validated by external baselines and ablations; only minor self-citations in TO/hardware tooling.

full rationale

The derivation chain is self-contained rather than circular. TO-generated 2D data (Eqs. 1-2) are produced by the paper's own impulse-scale optimization; TVAE and gait decoders are trained with reconstruction losses; the RL policy is optimized in 3D Isaac Gym with PPO and uses the frozen decoder only as a prior (τ_input = τ_dec + k_p(...) - k_d(·)); and the deployed student is distilled from a teacher with DAgger. None of these stages fits a parameter to the final real-world speed or success claims and then re-predicts it. The load-bearing ablations are genuine external checks: scaling τ_dec to 0 collapses success to 2.9% vs 94.6% at full scale (Fig. S8), and comparisons to AMP, vanilla RL, and HRL-with-residual are independent baselines. Self-citations appear for the 2D TO method (refs 13,15) and for the KAIST HOUND platform (ref 19), but both are tool/hardware inputs specified by in-text equations or prior platform design, not self-referential support that forces the claimed result. The supplement's algebraic PD-target equivalence (q_ref = a_scale·a_aux + τ_dec/k_p) is explicitly admitted as an identity, not a prediction, so it does not make the central claim definitional. The Statistical Analysis section states 'no statistical hypothesis tests, p-values, or confidence intervals were computed'; this limits confidence in the headline 4.25 and 6 m/s real-world measurements and in the zero-shot claim given the LiDAR preprocessing and vibration absorber, but it is an uncertainty/certainty issue rather than circularity. Overall, no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claims rest on transferability of 2D TO priors, simulator fidelity, and distillation reliability; no new physical entities, forces, or conserved quantities are introduced. The APT latent space is an internal learned representation with no independent falsifiable handle outside the paper.

free parameters (7)
  • TO periodicity cost weights λ1, λ2 = λ1=0.5; λ2=1.0 (bound), 0.0 (trot)
    Hand-set in Eq. 2; controls pitch regularization and thus the diversity/quality of the generated TO motion dataset.
  • PD gains kp, kd and auxiliary action scale ascale = kp=80, kd=2, ascale=0.2
    Final torque law τ_input=τ_dec+kp(q_default−q_t+ascale·a_aux)−kd·qdot; hand-tuned; central to real-robot stability.
  • TVAE latent dimension and KL weight = dim(z)=16, λ_KL=0.1
    Representation capacity; hand-chosen; affects the latent action space the RL policy explores.
  • RL exploration hyperparameters = entropy coef 0.001; latent KL coef 2.5e-6; PPO lr 3e-4; batch 204800
    Hand-set; modulate exploration/exploitation and sample efficiency.
  • Reward weights in Table S3 = See Table S3
    Hand-tuned task/regularization/gait/style reward weights; shape natural conservative obstacle negotiation; not fitted by a formal procedure.
  • Gait-selection period = 2 Hz (0.5 s hold)
    Assumes that at 4-7 m/s the robot needs >4 m lookahead for at least two decision cycles; hand-set.
  • Curriculum difficulty schedule = 10 levels + 2 flat levels; terrain parameter ranges in Supp.
    Hand-designed curriculum; affects the trained policy's robustness distribution.
axioms (6)
  • domain assumption A 2D sagittal-plane single rigid-body model with Bézier ground-reaction forces captures salient full-body dynamics for quadruped locomotion.
    Invoked in 'Impulse scale-based trajectory optimization'; the whole dataset is generated under this simplification.
  • standard math Momentum conservation over the gait cycle (Eq. 1) is sufficient to produce physically valid periodic trot/bound trajectories.
    Used to justify the periodicity conditions and reduce optimization complexity in Eq. 2.
  • domain assumption Isaac Gym's simulated contacts, actuator models, and terrain interactions are faithful enough for zero-shot sim-to-real transfer without explicit domain randomization.
    The policy is trained in Isaac Gym and deployed directly; no domain randomization or real-world fine-tuning is described.
  • domain assumption Teacher-student distillation via DAgger+BPTT transfers the privileged exteroceptive encoder to a student operating on real depth images and 2D LiDAR.
    Used in 'Distillation of exteroception latent using teacher-student framework'; deployment is stated as zero-shot.
  • ad hoc to paper The pretrained flat-ground 2D torque decoders remain a non-redundant action prior for 3D obstacle skills, with the auxiliary action covering only residual effects.
    Central to APT-RL; supported by Fig. S8 ablation, but the transfer from 2D flat TO data to 3D terrain is an assumption the paper does not derive.
  • domain assumption Gait switching at 2 Hz with a 0.5 s hold and a boolean trigger is sufficient for stable transitions at speeds up to 7 m/s.
    Introduced in the RL action section; the paper provides no ablation over switching frequency.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Agile perceptive multi-skill locomotion for quadrupedal robots in the wild." pith.science (2026). https://pith.science/paper/WOKFTDVY

@misc{pith2026260713579,
  author       = {Pith},
  title        = {Pith review of: Agile perceptive multi-skill locomotion for quadrupedal robots in the wild},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOKFTDVY}},
  note         = {Machine review of arXiv:2607.13579}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Enabling quadrupedal robots to traverse complex terrains-from rugged outdoor environments to urban landscapes-requires seamless integration of multiple motor skills, smooth transitions between gaits, and high-speed perceptive locomotion using only onboard sensors. We present APT-RL (Action Pretrained Transformer-based Reinforcement Learning), a unified framework that enables multi-skill locomotion to achieve high-speed traversal in complex environments through autonomous skill transitions utilizing only onboard perception and computation. Our approach generates large-scale, feature-rich 2D motion datasets through trajectory optimization with simplified dynamics. These datasets enable training of diverse, reusable locomotion skills that transfer effectively to a real quadruped robot operating on complex uneven terrains. The resulting high-quality skills serve as strong priors for efficient learning of complex downstream tasks and extend naturally to 3D environments, enabling smooth, high-speed multi-skill locomotion in deployed policy. Real-world experiments demonstrate the framework's capabilities: the robot performs agile maneuvers through complex indoor obstacles and outdoor wild environments, including dynamic drop-down maneuvers that reach instantaneous peak speeds of up to 6 meters per second. A single onboard policy enables robust traversal of diverse obstacles, including stairs, hurdles, stepping stones, gaps, and fallen branches, demonstrating the versatility and effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2607.13579 by Hae-Won Park, Jaehyun Park, Joon-ha Kim, Jun-Gill Kang, Seungwoo Hong, Tae-Gyu Song.

Figure 1
Figure 1. Figure 1: Traversing challenging terrains. (A) Our controller enables fast perceptive gait and skill selection to enable high-speed obstacle traversal. (B) The same controller was used to deploy over various outdoor terrains. with substantially reduced model dimensionality and complex- ity. This simplification not only reduces computational overhead [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training pipeline for APT-RL. (A) Our approach consists of motion dataset generation, representation learning, and rein￾forcement learning. Using trajectory optimization with simplified dynamics, we generate large-scale motion data to train diverse locomotion skills. These serve as action priors for efficient downstream learning and enable autonomous locomotion in complex terrains. (B) (i) Learning a share… view at source ↗
Figure 3
Figure 3. Figure 3: Agile perceptive locomotion with adaptive skills in various scenarios. (A) Urban scenario, (B) Wild scenario, and (C) Indoor scenario. Cmd vx is the commanded forward velocity, which was visualized through color changes along the trajectory. Shaded blue regions indicate periods when the trotting gait was selected, whereas red indicates bounding. Gait Select. denotes the gait selection output, where values … view at source ↗
Figure 4
Figure 4. Figure 4: Effectiveness of gaits under different terrains. Performance comparison of five gait decoders, trot, bound, pace, gallop, and pronk, across multiple terrain types. For the stair terrain, we used a lower command velocity range (train: -1 to 2m/s, test: 1 to 2m/s), whereas the other terrains were tested under a higher command velocity range (train: -1 to 7m/s, test: 1 to 7m/s). Each policy was trained with t… view at source ↗
Figure 5
Figure 5. Figure 5: Latent action embeddings and auxiliary torque analysis. (A) PCA embeddings of latent actions for trotting and bound￾ing across pretraining, RL, and real-world deployment. (B) t-SNE embeddings of real-world latent actions on flat, stepping stones, stairs, high-step, and log terrains. (C–E) Representative real-world behaviors (log jump, leg fracture during walking, in-place rota￾tion) together with their cor… view at source ↗
Figure 6
Figure 6. Figure 6: Autonomous gaits and motor skills selection. (A) Our robot uses different gaits and motor skills to traverse obstacles, depending on its state and command. (B) Gait fractions over continuous command velocities for various terrains and difficulty levels. (C) Task-performance metrics for automatic and fixed gait strategies. (i) Success rate, velocity tracking, and 1/COT across three command-speed groups (Low… view at source ↗
Figure 7
Figure 7. Figure 7: Evaluation across training strategies and baselines. (A) Training terrains used across all experiments. (B) Comparison be￾tween our method, AMP, and Vanilla RL. (C) Comparison of training curves on full obstacle terrain between our method and HRL with a residual. (D) Comparison of sample consumption between our method and HRL with a residual baseline in full obstacle terrain. (E) Evaluation of dynamic gait… view at source ↗
Figure 8
Figure 8. Figure 8: Sensor Ablation Study. Success rate comparison of depth and LiDAR, LiDAR-only, and depth-only policies across diverse terrains. For each experiment, 100 agents were evaluated per trial over 10 trials, and the error bars represent the standard deviation. Impulse scale-based trajectory optimization Unlike prior methods that rely solely on motion data from ani￾mal or human demonstrations [53, 63, 65], our app… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

84 extracted references · 8 linked inside Pith

  1. [1]

    D. F . Hoyt, C. R. Taylor, Gait and the energetics of locomotion in horses. Nature292(5820), 239–240 (1981)

    1981

  2. [2]

    A. N. Wimberly, G. J. Slater, M. C. Granatosky, Evolutionary history of quadrupedal walking gaits shows mammalian release from locomotor constraint.Proceedings of the Royal Society B288(1957), 20210937 (2021)

    1957

  3. [3]

    Martin, C

    C. Martin, C. Diederich, F . Verheggen, Cadaver Dogs and the Deathly Hallows—A Survey and literature review on selection and training procedure.Animals10(7), 1219 (2020)

    2020

  4. [4]

    Fenton, The use of dogs in search, rescue and recovery.Journal of Wilderness Medicine3(3), 292–300 (1992)

    V. Fenton, The use of dogs in search, rescue and recovery.Journal of Wilderness Medicine3(3), 292–300 (1992)

    1992

  5. [5]

    K. E. Jones, K. Dashfield, A. B. Downend, C. M. Otto, Search-and- rescue dogs: an overview for veterinarians.Journal of the American Veterinary Medical Association225(6), 854–860 (2004)

    2004

  6. [6]

    A. J. Ijspeert, Central pattern generators for locomotion control in animals and robots: a review.Neural networks21(4), 642–653 (2008)

    2008

  7. [7]

    S. A. Burden, T. Libby, K. Jayaram, S. Sponberg, J. M. Donelan, Why animals can outrun robots.Science Robotics9(89), eadi9754 (2024)

    2024

  8. [8]

    Kalakrishnan, J

    M. Kalakrishnan, J. Buchli, P . Pastor, M. Mistry, S. Schaal, Fast, robust quadruped locomotion over challenging terrain, inIEEE International Conference on Robotics and Automation(IEEE) (2010), pp. 2665– 2670

    2010

  9. [9]

    Semini, N

    C. Semini, N. G. Tsagarakis, E. Guglielmino, M. Focchi, F . Cannella, D. G. Caldwell, Design of HyQ–a hydraulically and electrically actu- ated quadruped robot.Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering225(6), 831–849 (2011)

    2011

  10. [10]

    Hutter, C

    M. Hutter, C. Gehring, M. Bloesch, M. A. Hoepflinger, C. D. Remy, R. Siegwart, StarlETH: A compliant quadrupedal robot for fast, efficient, and versatile locomotion, inAdaptive mobile robotics(World Scientific), pp. 483–490 (2012)

    2012

  11. [11]

    D. J. Hyun, S. Seok, J. Lee, S. Kim, High speed trot-running: Imple- mentation of a hierarchical controller using proprioceptive impedance control on the MIT Cheetah.The International Journal of Robotics Research33(11), 1417–1445 (2014)

    2014

  12. [12]

    H.-W. Park, P . M. Wensing, S. Kim, Online Planning for Autonomous Running Jumps Over Obstacles in High-Speed Quadrupeds, in Robotics: Science and Systems(RSS Foundation) (2015)

    2015

  13. [13]

    H.-W. Park, S. Park, S. Kim, Variable-speed quadrupedal bounding using impulse planning: Untethered high-speed 3d running of MIT Cheetah 2, inIEEE international conference on Robotics and automa- tion (ICRA)(IEEE) (2015), pp. 5163–5170

    2015

  14. [14]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V. Tsou- nis, J. Hwangbo, K. Bodie, P . Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, M. Hoepflinger, ANYmal-a highly mobile and dynamic quadrupedal robot, inIEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE) (2016), pp. 38–44

    2016

  15. [15]

    H.-W. Park, P . M. Wensing, S. Kim, High-speed bounding with the MIT Cheetah 2: Control design and experiments.The International Journal of Robotics Research36(2), 167–192 (2017)

    2017

  16. [16]

    Hong, J.-H

    S. Hong, J.-H. Kim, H.-W. Park, Real-time constrained nonlinear model predictive control on SO (3) for dynamic legged locomotion, inIEEE/RSJ International Conference on Intelligent Robots and Sys- tems (IROS)(IEEE) (2020), pp. 3982–3989

    2020

  17. [17]

    Y . Ding, A. Pandala, C. Li, Y .-H. Shin, H.-W. Park, Representation-free model predictive control for dynamic motions in quadrupeds.IEEE Transactions on Robotics37(4), 1154–1171 (2021)

    2021

  18. [18]

    S. Hong, Y . Um, J. Park, H.-W. Park, Agile and versatile climbing on ferromagnetic surfaces with a quadrupedal robot.Science Robotics 7(73), eadd1017 (2022)

    2022

  19. [19]

    Y .-H. Shin, S. Hong, S. Woo, J. Choe, H. Son, G. Kim, J.-H. Kim, K. Lee, J. Hwangbo, H.-W. Park, Design of KAIST HOUND, a quadruped robot platform for fast and efficient locomotion with mixed-integer nonlinear optimization of a gear train, inInternational Conference on Robotics and Automation (ICRA)(IEEE) (2022), pp. 6614–6620

    2022

  20. [20]

    Grandia, F

    R. Grandia, F . Jenelten, S. Y ang, F . Farshidian, M. Hutter, Perceptive locomotion through nonlinear model-predictive control.IEEE Transac- tions on Robotics39(5), 3402–3421 (2023)

    2023

  21. [21]

    G. Kim, D. Kang, J.-H. Kim, S. Hong, H.-W. Park, Contact-implicit Model Predictive Control: Controlling diverse quadruped motions without pre- planned contact modes or trajectories.The International Journal of Robotics Research44(3), 486–510 (2025)

    2025

  22. [22]

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, V. Vanhoucke, Sim-to-real: Learning agile locomotion for quadruped robots, inRobotics: Science and Systems(RSS Foundation) (2018)

    2018

  23. [23]

    Iscen, K

    A. Iscen, K. Caluwaerts, J. Tan, T. Zhang, E. Coumans, V. Sindhwani, V. Vanhoucke, Policies modulating trajectory generators, inConference on Robot Learning(PMLR) (2018), pp. 916–926

    2018

  24. [24]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots. Science Robotics4(26), eaau5872 (2019)

    2019

  25. [25]

    Shin, T.-G

    Y .-H. Shin, T.-G. Song, G. Ji, H.-W. Park, Reinforcement Learning for High-Speed Quadrupedal Locomotion With Motor Operating Region Constraints: Mitigating Motor Model Discrepancies through Torque Clip- ping in Realistic Motor Operating Region.IEEE Robotics & Automation Magazine(2024)

    2024

  26. [26]

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

    2020

  27. [27]

    G. Ji, J. Mun, H. Kim, J. Hwangbo, Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion. IEEE Robotics and Automation Letters7(2), 4630–4637 (2022)

    2022

  28. [28]

    S. Choi, G. Ji, J. Park, H. Kim, J. Mun, J. H. Lee, J. Hwangbo, Learn- ing quadrupedal locomotion on deformable terrain.Science Robotics 8(74), eade2256 (2023)

    2023

  29. [29]

    I. M. A. Nahrendra, B. Yu, H. Myung, DreamWaQ: Learning robust quadrupedal locomotion with implicit terrain imagination via deep rein- forcement learning, inIEEE International Conference on Robotics and Automation (ICRA)(IEEE) (2023), pp. 5078–5084

    2023

  30. [30]

    S. Chen, B. Zhang, M. W. Mueller, A. Rai, K. Sreenath, Learning torque control for quadrupedal locomotion, inIEEE-RAS International Conference on Humanoid Robots (Humanoids)(IEEE) (2023), pp. 1–8

    2023

  31. [31]

    Kim, Y .-H

    G. Kim, Y .-H. Lee, H.-W. Park, A Learning Framework for Diverse Legged Robot Locomotion Using Barrier-Based Style Rewards.arXiv preprint arXiv:2409.15780(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    G. B. Margolis, G. Y ang, K. Paigwar, T. Chen, P . Agrawal, Rapid locomo- tion via reinforcement learning.The International Journal of Robotics Research43(4), 572–587 (2024)

    2024

  33. [33]

    Y . Kim, H. Oh, J. Lee, J. Choi, G. Ji, M. Jung, D. Y oum, J. Hwangbo, Not Only Rewards but Also Constraints: Applications on Legged Robot Locomotion.IEEE Transactions on Robotics40, 2984–3003 (2024)

    2024

  34. [34]

    W. Yu, J. Tan, C. K. Liu, G. Turk, Preparing for the unknown: Learning a universal policy with online system identification, inRobotics: Science and Systems(RSS Foundation) (2017)

    2017

  35. [35]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, J. Malik, RMA: Rapid Motor Adaptation for Legged Robots, inRobotics: Science and Systems(RSS Foundation) (2021)

    2021

  36. [36]

    Rudin, D

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

    2022

  37. [37]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hutter, Learn- ing robust perceptive locomotion for quadrupedal robots in the wild. Science robotics7(62), eabk2822 (2022)

    2022

  38. [38]

    T. Miki, J. Lee, L. Wellhausen, M. Hutter, Learning to walk in con- fined spaces using 3d representation.arXiv preprint arXiv:2403.00187 (2024)

    Pith/arXiv arXiv 2024

  39. [39]

    S. Luo, S. Li, R. Yu, Z. Wang, J. Wu, Q. Zhu, PIE: Parkour With Implicit- Explicit Learning Framework for Legged Robots.IEEE Robotics and Automation Letters9(11), 9986–9993 (2024)

    2024

  40. [40]

    Made Aswin Nahrendra, B

    I. Made Aswin Nahrendra, B. Yu, M. Oh, D. Lee, S. Lee, H. Lee, H. Lim, H. Myung, Obstacle-Aware Quadrupedal Locomotion With Resilient Multi-Modal Reinforcement Learning.arXiv e-printspp. arXiv–2409 (2024)

    2024

  41. [41]

    Bellegarda, A

    G. Bellegarda, A. Ijspeert, CPG-RL: Learning central pattern genera- tors for quadruped locomotion.IEEE Robotics and Automation Letters Research Article 18 7(4), 12547–12554 (2022)

    2022

  42. [42]

    Y . Jin, X. Liu, Y . Shao, H. Wang, W. Y ang, High-speed quadrupedal locomotion by imitation-relaxation reinforcement learning.Nature Ma- chine Intelligence4(12), 1198–1208 (2022)

    2022

  43. [43]

    Zhuang, Z

    Z. Zhuang, Z. Fu, J. Wang, C. G. Atkeson, S. Schwertfeger, C. Finn, H. Zhao, Robot Parkour Learning, inConference on Robot Learning (CoRL)(2023)

    2023

  44. [44]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, D. Pathak, Extreme parkour with legged robots, inIEEE International Conference on Robotics and Automation (ICRA)(2024)

    2024

  45. [45]

    Hoeller, N

    D. Hoeller, N. Rudin, D. Sako, M. Hutter, ANYmal Parkour: Learn- ing agile navigation for quadrupedal robots.Science Robotics9(88), eadi7566 (2024)

    2024

  46. [46]

    H. Kim, H. Oh, J. Park, Y . Kim, D. Y oum, M. Jung, M. Lee, J. Hwangbo, High-speed control and navigation for quadrupedal robots on complex and discrete terrain.Science Robotics10(102), eads6192 (2025)

    2025

  47. [47]

    X. B. Peng, Z. Ma, P . Abbeel, S. Levine, A. Kanazawa, AMP: Adver- sarial motion priors for stylized physics-based character control.ACM Transactions on Graphics (ToG)40(4), 1–20 (2021)

    2021

  48. [48]

    Escontrela, X

    A. Escontrela, X. B. Peng, W. Yu, T. Zhang, A. Iscen, K. Goldberg, P . Abbeel, Adversarial motion priors make good substitutes for complex reward functions, inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 25–32

    2022

  49. [49]

    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, Z. Zhang, Lifelike agility and play in quadrupedal robots using reinforcement learning and generative pre-trained models. Nature Machine Intelligencepp. 1–12 (2024)

    2024

  50. [50]

    Brakel, S

    P . Brakel, S. Bohez, L. Hasenclever, N. Heess, K. Bousmalis, Learn- ing coordinated terrain-adaptive locomotion by imitating a centroidal dynamics planner, inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 10335–10342

    2022

  51. [51]

    Fuchioka, Z

    Y . Fuchioka, Z. Xie, M. Van de Panne, OPT -Mimic: Imitation of opti- mized trajectories for dynamic quadruped behaviors, inIEEE Interna- tional Conference on Robotics and Automation (ICRA)(IEEE) (2023), pp. 5092–5098

    2023

  52. [52]

    D. Kang, J. Cheng, M. Zamora, F . Zargarbashi, S. Coros, Rl+ Model- based control: Using on-demand optimal control to learn versatile legged locomotion.IEEE Robotics and Automation Letters8(10), 6619–6626 (2023)

    2023

  53. [53]

    J. Wu, G. Xin, C. Qi, Y . Xue, Learning robust and agile legged locomo- tion using adversarial motion priors.IEEE Robotics and Automation Letters8(8), 4975–4982 (2023)

    2023

  54. [54]

    Jenelten, J

    F . Jenelten, J. He, F . Farshidian, M. Hutter, DTC: Deep Tracking Control. Science Robotics9(86), eadh5401 (2024)

    2024

  55. [55]

    A. G. Barto, S. Mahadevan, Recent advances in hierarchical reinforce- ment learning.Discrete event dynamic systems13, 341–379 (2003)

    2003

  56. [56]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, J. Dean, Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer, inInternational Conference on Learning Representations(2017)

    2017

  57. [57]

    Y ang, K

    C. Y ang, K. Yuan, Q. Zhu, W. Yu, Z. Li, Multi-expert learning of adaptive legged locomotion.Science Robotics5(49), eabb2174 (2020)

    2020

  58. [58]

    Shafiee, G

    M. Shafiee, G. Bellegarda, A. Ijspeert, Viability leads to the emer- gence of gait transitions in learning agile quadrupedal locomotion on challenging terrains.Nature Communications15(1), 3073 (2024)

    2024

  59. [59]

    Bellegarda, M

    G. Bellegarda, M. Shafiee, A. Ijspeert, Allgaits: Learning all quadruped gaits and transitions, inInternational Conference on Robotics and Automation (ICRA)(IEEE) (2025), pp. 15929–15935

    2025

  60. [60]

    Owaki, A

    D. Owaki, A. Ishiguro, A quadruped robot exhibiting spontaneous gait transitions from walking to trotting to galloping.Scientific reports7(1), 277 (2017)

    2017

  61. [61]

    Hasenclever, F

    L. Hasenclever, F . Pardo, R. Hadsell, N. Heess, J. Merel, CoMic: Com- plementary task learning & mimicry for reusable skills, inInternational Conference on Machine Learning(PMLR) (2020), pp. 4105–4115

    2020

  62. [62]

    S. Liu, G. Lever, Z. Wang, J. Merel, S. M. A. Eslami, D. Hennes, W. M. Czarnecki, Y . Tassa, S. Omidshafiei, A. Abdolmaleki, N. Y . Siegel, L. Hasenclever, L. Marris, S. Tunyasuvunakool, H. F . Song, M. Wulfmeier, P . Muller, T. Haarnoja, B. Tracey, K. Tuyls, T. Graepel, N. Heess, From motor control to team play in simulated humanoid football.Science Robo...

    2022

  63. [63]

    X. B. Peng, Y . Guo, L. Halper, S. Levine, S. Fidler, ASE: Large-scale reusable adversarial skill embeddings for physically simulated charac- ters.ACM Transactions On Graphics (TOG)41(4), 1–17 (2022)

    2022

  64. [64]

    Z. Luo, J. Cao, J. Merel, A. Winkler, J. Huang, K. M. Kitani, W. Xu, Uni- versal Humanoid Motion Representations for Physics-Based Control, inInternational Conference on Learning Representations(2023)

    2023

  65. [65]

    Bohez, S

    S. Bohez, S. Tunyasuvunakool, P . Brakel, F . Sadeghi, L. Hasenclever, Y . Tassa, E. Parisotto, J. Humplik, T. Haarnoja, R. Hafner, M. Wulfmeier, M. Neunert, B. Moran, N. Siegel, A. Huber, F . Romano, N. Batchelor, F . Casarini, J. Merel, R. Hadsell, N. Heess, Imitate and repurpose: Learning reusable robot movement skills from human and animal be- haviors...

    Pith/arXiv arXiv 2022

  66. [66]

    A. L. Mitchell, W. X. Merkt, M. Geisert, S. Gangapurwala, M. Engelcke, O. P . Jones, I. Havoutis, I. Posner, VAE-Loco: Versatile quadruped locomotion by learning a disentangled gait representation.IEEE Trans- actions on Robotics39(5), 3805–3820 (2023)

    2023

  67. [67]

    J. Wu, Y . Xue, C. Qi, Learning multiple gaits within latent space for quadruped robots.arXiv preprint arXiv:2308.03014(2023)

    Pith/arXiv arXiv 2023

  68. [68]

    A. L. Mitchell, W. Merkt, A. Papatheodorou, I. Havoutis, I. Posner, Gaitor: Learning a Unified Representation Across Gaits for Real-World Quadruped Locomotion, inConference on Robot Learning(PMLR) (2024)

    2024

  69. [69]

    J.-G. Kang, D. Lee, S. Han, A Highly Maneuverable Flying Squirrel Drone with Controllable Foldable Wings, inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2023), pp. 6652–6659

    2023

  70. [70]

    Petrovich, M

    M. Petrovich, M. J. Black, G. Varol, Action-Conditioned 3D human mo- tion synthesis with transformer VAE, inProceedings of the IEEE/CVF International Conference on Computer Vision(2021), pp. 10985– 10995

    2021

  71. [71]

    Silver, K

    T. Silver, K. Allen, J. Tenenbaum, L. Kaelbling, Residual policy learning. arXiv preprint arXiv:1812.06298(2018)

    Pith/arXiv arXiv 2018

  72. [72]

    Johannink, S

    T. Johannink, S. Bahl, A. Nair, J. Luo, A. Kumar, M. Loskyll, J. A. Ojea, E. Solowjow, S. Levine, Residual reinforcement learning for robot control, inInternational Conference on Robotics and Automation (ICRA)(IEEE) (2019), pp. 6023–6029

    2019

  73. [73]

    Makoviychuk, L

    V. Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, G. State, Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470(2021)

    Pith/arXiv arXiv 2021

  74. [74]

    Schulman, F

    J. Schulman, F . Wolski, P . Dhariwal, A. Radford, O. Klimov, Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347(2017)

    Pith/arXiv arXiv 2017

  75. [75]

    Agarwal, A

    A. Agarwal, A. Kumar, J. Malik, D. Pathak, Legged locomotion in challenging terrains using egocentric vision, inConference on robot learning(PMLR) (2023), pp. 403–415

    2023

  76. [76]

    Kareer, N

    S. Kareer, N. Y okoyama, D. Batra, S. Ha, J. Truong, ViNL: Visual naviga- tion and locomotion over obstacles, inIEEE International Conference on Robotics and Automation (ICRA)(IEEE) (2023), pp. 2018–2024

    2023

  77. [77]

    Y ang, G

    R. Y ang, G. Y ang, X. Wang, Neural volumetric memory for visual locomotion control, inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition(2023), pp. 1430–1440

    2023

  78. [78]

    S. Ross, G. Gordon, J. A. Bagnell, A reduction of imitation learning and structured prediction to no-regret online learning, inProceedings of the International Conference on Artificial Intelligence and Statistics (JMLR Workshop and Conference Proceedings) (2011), pp. 627–635

    2011

  79. [79]

    Accessed 7 May 2026

    Morvélo Bicycle Apparel, Trail Dogs, Y ouTube video (2017), https: //www.youtube.com/watch?v=w5y6f3dOp00, licensed under a Creative Commons Attribution License (CC BY). Accessed 7 May 2026

    2017

  80. [80]

    youtube.com/watch?v=kqnB76cQSd0, licensed under a Creative Com- mons Attribution License (CC BY)

    FOX Sports, Check out the best of the 2019 WKC Masters Agility Championship | FOX SPORTS, Y ouTube video (2019), https://www. youtube.com/watch?v=kqnB76cQSd0, licensed under a Creative Com- mons Attribution License (CC BY). Accessed 7 May 2026

    2019

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.