Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A two-layer controller that imitates five real dog gaits and reads only depth images lets a quadrupedal robot run a randomized six-obstacle agility course at 1.1 m/s, reaching 3.2 m/s over hurdles.

desk verdict A real depth-only quadruped controller that switches learned gaits and completes an agility course on hardware; the headline success rate is real but statistically under-documented. read the letter →

arxiv 2505.09979 v1 pith:PXW5GPSY submitted 2025-05-15 cs.RO

Huiqiao Fu , Haoyu Dong , Wentao Xu , Zhehao Zhou , Guizhou Deng , Kaiqiang Tang , Daoyi Dong , Chunlin Chen This is my paper · ORCID
classification cs.RO
keywords quadrupedalroboticsimitationlearningsim-to-realtransferadversarialprivilegeddepthperceptionlocomotiondogagility
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 claims that a quadrupedal robot can achieve dog-like agility by splitting control into two learned layers: a Basic Behavior Controller that imitates diverse real-dog gaits, and a Task-Specific Controller that decides, from depth-camera images alone, which gait to use and how fast to go. On a Unitree Go2 robot, the authors report the first autonomous completion of a randomized six-obstacle agility course — A-frame, bar-jump, poles, seesaw, tire-jump, and tunnel — at an average speed of 1.1 m/s with a success rate above 78%, plus a peak speed of 3.2 m/s in a hurdling task. The decisive step is transfer: they tune simulator physics with about 80 seconds of real-world data, fine-tune the behavior controller inside that improved simulator, and argue the resulting behaviors transfer with minimal performance loss. If true, the paper shows that broad, natural, animal-like behavior libraries can be built from motion-capture data and shipped to real robots with cheap sensing, opening agility-style autonomy to commodity hardware.

What carries the argument

The argument rests on three components. The BBC is trained with semi-supervised InfoGAIL: a least-squares discriminator (the AMP-style reward) scores how closely transitions match real dog motion, while a predictor maximizes mutual information between latent variables and transitions, so the policy learns to produce specific gaits on command — the discrete skill latent $c$ chooses the mode, the continuous shift latent $\epsilon$ varies style within a mode, and RIM lets the latent distribution rebalance toward the true, imbalanced data. The TSC is a student policy trained by privileged learning: a teacher trained with Hybrid-PPO in a hybrid action space sees scandots, waypoint yaws, and obstacle type, and the student, which sees only depth images through a GRU encoder, learns to imitate it while a BYOL contrastive loss and probabilistic augmentations (white noise, blur, cropping) make the depth features robust. The third component, EASI, turns simulator identification into an adversarial search: an evolutionary strategy proposes distributions over joint PD gains, a neural discriminator scores transitions by realism, and the loop converges in under ten minutes on 80 seconds of real data, after which the BBC is fine-tuned in the enhanced simulator.

What would settle it

Run the full randomized agility challenge twenty times with the same controller and count per-trial successes: the claimed success rate above 78% predicts no more than four failures, and a systematic failure on any single obstacle would expose a gap in the behavior library rather than in perception. Separately, command motions outside the 80-second calibration set (sustained 3 m/s trotting, or sharp turning on a different floor surface) and compare real joint trajectories with simulator rollouts under EASI's identified parameters: divergence beyond the calibration-time spread would refute the minimal-loss transfer claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that quadrupedal agility decomposes into a reusable behavior library and a task coordinator, and that both can be learned and transferred. The Basic Behavior Controller is a single policy, trained by a semi-supervised variant of InfoGAIL, that reproduces five dog gaits — walk, pace, trot, canter, jump — from raw motion capture data with less than 5% of it labeled: a discrete latent variable selects the gait, a continuous latent variable selects style, and a Regularized Information Maximization term lets the policy learn from imbalanced unlabeled data. The Task-Specific Controller is a student policy trained by privileged learning to imitate a teacher that sees scandots, waypoint angles, and obstacle types; the student sees only a depth image, with a BYOL self-supervised loss and random augmentations shielding it from real-world image noise. The simulator is then brought closer to reality by EASI, an evolutionary search over motor parameters scored by a neural discriminator, using 80 seconds of real-world data and under ten minutes of GPU time. With these pieces, the robot switches between trot, canter, and jump by itself, jumps onto a 30 cm box, and clears hurdles at 3.2 m/s — evidence, the paper argues, that a substantial step toward animal-like agility has been made.

Load-bearing premise

The results depend on the simulator, after EASI tuning on about 80 seconds of real-world data, matching the real robot's dynamics closely enough that the fine-tuned BBC transfers with minimal performance loss; the paper's evidence for this match is qualitative — joint frequency spectra and trajectories — rather than a measured success-rate comparison in simulation versus reality.

Editorial extensions

If this is right

  • A quadruped with a single depth camera can navigate a randomized course of six visually distinct obstacles without global localization, so agility does not require external tracking or motion-capture infrastructure.
  • Because the BBC is task-agnostic, the same five-gait library can serve new downstream tasks by training only a new TSC; box jumping, hurdling, and human-collaborative routing are demonstrated instances of this reuse.
  • The reported success rate above 78% at 1.1 m/s on a commercial robot, against the baseline parkour controller, sets a concrete benchmark — the paper's own agility challenge — that other groups can reproduce and beat.
  • EASI's use of only 80 seconds of real data and under ten minutes of GPU search suggests that simulator tuning can replace hand-set domain randomization ranges as the standard sim-to-real step for legged robots.
  • The robot clears bar-jumps up to 0.35 m and tire-jumps up to 0.7 m by switching autonomously from canter to jump using depth alone, which the paper shows beats both the parkour baseline and a human teleoperator on the same course.

Reading between the lines

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

  • Because EASI accepts any controller for collecting calibration trajectories, the 80-second real-world step could be automated and repeated over the robot's lifetime, letting the simulator track actuator wear, battery voltage, or payload changes — a maintenance loop the paper does not explore.
  • The BBC's controllability via two latent variables suggests the same library could be driven by other high-level interfaces, including voice- or gesture-based human commands, which the paper lists as future work; a low-cost teleoperation test would be a direct way to verify this.
  • The student's reliance on a short depth history invites a concrete test of whether the reported success rate degrades with a narrower camera field of view or longer obstacle runs; if it does, the limiting factor is perception range rather than the gait library, redirecting effort toward wider-angle or multimodal sensing.
  • Nothing restricts the behavior library to dogs: retraining the BBC on motion capture from other quadrupeds or on synthetic gaits could probe whether the agility gains come from the specific five gaits themselves or from the semi-supervised training objective that disentangles them.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 an integrated controller for a Unitree Go2 quadruped, combining a Basic Behavior Controller (BBC) trained by a semi-supervised variant of InfoGAIL from dog motion capture data, a Task-Specific Controller (TSC) trained by privileged learning with depth images, and an Evolutionary Adversarial Simulator Identification (EASI) procedure for sim-to-real transfer. The authors report that the robot completes a randomized quadrupedal agility course at an average speed of 1.1 m/s, reaches 3.2 m/s during hurdling, achieves a success rate exceeding 78% on the full course, and exhibits five natural gait modes with continuous style variation, with ablations for the BBC, TSC, and EASI components.

Significance. If the empirical claims hold, this is a meaningful advance at the intersection of imitation learning, hierarchical control, and sim-to-real transfer: it demonstrates a single integrated controller that switches among diverse natural behaviors in real time using depth perception alone. The paper's explicit strengths include thorough ablations of the BBC (NMI/ENT metrics and t-SNE visualizations), comparisons against a strong parkour baseline, real-robot deployment, and a sim-to-sim validation of the EASI parameter search. The reported training times and data requirements are also concrete and reproducible in spirit. However, the central empirical claims are currently supported by a single 26 s run and by success-rate values with no trial counts, error bars, or failure definitions, and one of the formal objectives in the EASI derivation appears mis-specified. The overall approach is credible and potentially important, but the evidence base needs substantial strengthening before the headline claims can be accepted.

major comments (4)
  1. [Section 2.1 and Section 2.4] The headline claim that the robot 'successfully completes the quadrupedal agility challenge' with a success rate that 'exceeds 78%' is not statistically supported. The full-course result is presented as a single 26 s run (Fig. 2b), with no number of randomized trials, per-trial outcomes, confidence interval, or definition of failure (e.g., whether obstacle knock-down, falls, timeouts, or operator resets count as failures). In addition, Section 2.4 reports a simulated noise-robustness success rate of approximately 78% for the TSC, so it is unclear whether the real-world number is an independent measurement or a transferred simulation value. The authors should report trial counts and the full evaluation protocol, and present all success rates with error bars.
  2. [Section 4.4, Eq. (13)] Equation (13), as written, does not define the adversarial game described in the text. In a standard LSGAN formulation, the discriminator minimizes E_fake[(D+1)^2] + E_real[(D-1)^2] while the generator should minimize E_fake[(D-1)^2] (i.e., push fake transitions toward the real label). The paper instead states max_Ξ min_~D of that full expression, which would make the generator maximize the fake term E[(D+1)^2] and would drive the outer objective toward a degenerate solution because the inner minimization over ~D can set D(fake)=-1 independent of Ξ. This appears to be more than a notational slip: the sign and the min-max order need to be corrected or clarified, and the actual objective used in Algorithm S2 should be stated explicitly.
  3. [Section 2.3] The NMI/ENT comparison of behavioral diversity depends on a 'pre-trained behavior classifier' whose training labels, architecture, and accuracy are not specified. If this classifier is trained on the same five human-defined categories that the latent skill variable c is designed to encode, then the reported NMI/ENT values primarily demonstrate that the policy can be controlled by the latent code, rather than that the method discovers behavior modes from raw data. The paper should document the classifier and provide additional, independent validation of the naturalness and separability of the generated behaviors (e.g., human evaluation or ground-truth label metrics), so that the diversity claim is not circular.
  4. [Section 2.5 and Section 4.4] The evidence for the claim that the BBC transfers to the real world 'with minimal performance loss' after EASI fine-tuning is qualitative, consisting of joint frequency spectra and trajectories in Fig. 6c-d. The sim-to-sim experiment in Fig. 6a shows convergence to true parameters, but the sim-to-real experiment in Fig. 6b lacks quantitative transfer metrics such as success rate, tracking error, or task reward measured on the real robot. Since EASI is a core component of the proposed pipeline, the paper should provide quantitative sim-to-real transfer results with trial counts rather than relying on visual similarity of signals.
minor comments (5)
  1. [Section 2.1] Please clearly define the success/failure criteria for the agility challenge before the first use of the 'success rate' phrase, and state explicitly whether the 26 s completion time and the 1.1 m/s average speed come from one run or from an average over multiple runs.
  2. [General] The manuscript contains several typos: 'IssacGym' in Section 4.1 should be 'IsaacGym'; 'the the sim-to-real gap' in Section 1 has a duplicated article; and 'The results shows' in Section 2.3 should be 'The results show.'
  3. [Section 2.1 and Section 2.4] The text uses both 'exceeds 78%' and 'approximately 78%' for the success-rate claim; please disambiguate the real-world and simulated measurements and avoid carrying the same numerical threshold across different experiments.
  4. [Section 4.3, Eq. (3)] In Equation (3), the notation d_E^L(o_I|y) is not defined before use; please specify the conditional distribution of the observation window given the label for the labeled expert demonstrations.
  5. [General] The paper should include a data and code availability statement, or at least a permanent repository link, so that the reported trial counts and learned policies can be audited by the community.

Circularity Check

1 steps flagged · score 2.0 of 10

Low circularity: the agility and sim-to-real claims are independent; only the BBC diversity metric partly re-uses its own training labels.

  1. self definitional [Section 2.3 (Fig. 4g) and Section 4.3.1 (Eqs. 3-6)]
    "We use the latent skill variable c to represent five behavioral modes: walk, pace, trot, canter, and jump, and seek to maximize the semi-supervised mutual information I(c;oI). ... The category information in ENT and NMI is output by a pre-trained behavior classifier."

    The five 'discovered' modes are fixed in advance as the semantic content of c, and a small labeled subset (<5%) is used to train the predictor Q_c; the policy is rewarded with rSS = log Q_c(c|oI), i.e., it is explicitly trained to make state transitions predictable as one of these five labels. The ENT/NMI evaluation then scores how well the policy's c-conditioned clusters match a pre-trained behavior classifier whose categories are the same five human-defined modes. The high NMI/low ENT is therefore not evidence that five modes emerged from the raw data; it is the injected label signal returned by an evaluation metric whose ground truth is the same label set.

full rationale

The central derivation chain is not circular. The BBC is trained from real dog motion-capture data using a semi-supervised InfoGAIL variant (with less than 5% labels), the TSC is trained by privileged reinforcement learning with depth-image student distillation, and the EASI simulator identification uses 80 seconds of independently collected real-world trajectories; none of these stages defines its output in terms of its own conclusion. The real-robot agility challenge result is an external, falsifiable outcome, and the EASI tuning is evaluated in the paper with new joint-spectra and trajectory comparisons rather than merely cited. The only self-referential element is the ENT/NMI diversity evaluation: the ground-truth categories come from a pre-trained behavior classifier that encodes the same five human labels used to supervise the latent skill variable c, and the training reward rSS = log Q_c(c|oI) directly optimizes the same class-predictability that NMI measures. This makes the framing that the five modes are 'extracted' from raw data partly definitional, but it is a component-level evaluation bias, not the paper's central claim. The frequently noted absence of trial counts and error bars for the reported 'exceeds 78%' real-world success rate is a statistical reporting matter, not a circularity of the derivation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest primarily on the quality of the dog motion capture data, the tightness of the variational mutual-information bounds, the validity of the adversarial imitation signal, and the sufficiency of the privileged/distilled information for the TSC. The EASI procedure introduces a fitted set of simulator parameters, and several reward weights and command ranges are hand-selected. The paper introduces no new physical entities, forces, or conserved quantities.

free parameters (3)
  • Simulator physical parameters identified by EASI = not disclosed numerically (optimized over ~50 generations)
    EASI searches over joint PD gains, friction, and mass-related parameters (Section 4.4, Fig.6) using 80 seconds of real-robot data; the resulting parameter distribution is used to fine-tune the BBC, so the real-world behavior depends on these fitted values.
  • Reward weights w_I and w_T in the BBC total reward = w_I = 0.2, w_T = 0.2 (Table S1)
    These hand-set weights balance imitation and command-following rewards (Eq. 11); the learned behavior and its controllability depend on this balance, which is not derived from data.
  • Per-mode command ranges (vx, vy, omega_z, jump/locomotion height) = e.g., trot vx [0.5, 1.5] m/s, canter [0.8, 2.0] m/s (Table S2)
    The allowed ranges for velocity and height commands are chosen by hand for each gait; they define the envelope of behavior the BBC must cover and are used to assign task rewards in TSC training.
assumptions (4)
  • domain assumption The dog motion capture dataset [57] is a valid source of natural quadrupedal behaviors for the Go2 robot after retargeting.
    The BBC is trained entirely from this dataset (Section 4.3.5); if the retargeting or data relevance fails, the learned behaviors will not be natural or controllable on the Go2.
  • ad hoc to paper Optimizing the variational lower bounds in Eqs. (3)-(7) with a shared predictor Qc yields a policy that actually encodes the intended discrete and continuous latent variables.
    The paper relies on the InfoGAIL-style bound being sufficiently tight; this is an empirical claim supported only by the NMI/ENT metrics and t-SNE visualizations, not by a formal guarantee.
  • standard math The LSGAN discriminator (Eq. 1) provides a valid adversarial signal whose optimum minimizes the Pearson chi-square divergence between policy and expert state transitions.
    This is a known result from LSGAN (ref [53]) and is used in Eq. (2) to define the imitation reward.
  • domain assumption Privileged information (scandots, waypoints, obstacle type) is sufficient for a teacher policy to solve the agility task in simulation, and depth images plus proprioception are sufficient for the student to imitate the teacher.
    The TSC training (Section 4.5) assumes this two-stage distillation is valid; if the teacher's privileged observations omit critical state or the student cannot infer them from depth, the real-world controller would fail.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots." pith.science (2026). https://pith.science/paper/PXW5GPSY

@misc{pith2026250509979,
  author       = {Pith},
  title        = {Pith review of: Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXW5GPSY}},
  note         = {Machine review of arXiv:2505.09979}
}
read the original abstract

Achieving animal-like agility is a longstanding goal in quadrupedal robotics. While recent studies have successfully demonstrated imitation of specific behaviors, enabling robots to replicate a broader range of natural behaviors in real-world environments remains an open challenge. Here we propose an integrated controller comprising a Basic Behavior Controller (BBC) and a Task-Specific Controller (TSC) which can effectively learn diverse natural quadrupedal behaviors in an enhanced simulator and efficiently transfer them to the real world. Specifically, the BBC is trained using a novel semi-supervised generative adversarial imitation learning algorithm to extract diverse behavioral styles from raw motion capture data of real dogs, enabling smooth behavior transitions by adjusting discrete and continuous latent variable inputs. The TSC, trained via privileged learning with depth images as input, coordinates the BBC to efficiently perform various tasks. Additionally, we employ evolutionary adversarial simulator identification to optimize the simulator, aligning it closely with reality. After training, the robot exhibits diverse natural behaviors, successfully completing the quadrupedal agility challenge at an average speed of 1.1 m/s and achieving a peak speed of 3.2 m/s during hurdling. This work represents a substantial step toward animal-like agility in quadrupedal robots, opening avenues for their deployment in increasingly complex real-world environments.

Figures

Figures reproduced from arXiv: 2505.09979 by the authors.

Figure 7
Figure 7. In the first stage, a general-purpose BBC is trained using a semi-supervised variant of InfoGAIL to replicate common behaviors observed in real dog motion capture data. The latent skill variable c is learned in a semi-supervised manner and encodes five discrete behavior modes: walk, pace, trot, canter, and jump. Meanwhile, the latent shifting variable ϵ, learned unsupervisedly, captures continuous behavior styles. T… view at source ↗

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. Spatiotemporal Agility: Time-Constrained Reinforcement Learning for Vision-Guided Dynamic Quadrupedal Interception

    cs.RO 2026-08 conditional novelty 6.0 of 10

    A time-aware, position-conditioned reinforcement learning policy enables a quadruped to intercept thrown balls with higher success than a velocity-tracking baseline in the tested setups.

Reference graph

Works this paper leans on

62 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [1]

    Crufts.THE KENNEL CLUBhttps://www.crufts.org.uk/ (2024)

    work page 2024

  2. [2]

    Arm, P.et al.Scientific exploration of challenging planetary analog environments with a team of legged robots.Science Robotics8, eade9548 (2023)

    work page 2023

  3. [3]

    Miki, T.et al.Learning robust perceptive locomotion for quadrupedal robots in the wild.Science Robotics7, eabk2822 (2022)

    work page 2022

  4. [4]

    & El Manira, A

    Grillner, S. & El Manira, A. Current principles of motor control, with special reference to vertebrate locomotion.Physiological Reviews100, 271–320 (2020)

    work page 2020

  5. [5]

    M., Katz, B., Bledt, G

    Di Carlo, J., Wensing, P. M., Katz, B., Bledt, G. & Kim, S.Dynamic locomotion in the mit cheetah 3 through convex model-predictive control, 1–9 (IEEE, 2018)

    work page 2018

  6. [6]

    D., Jenelten, F., Gehring, C

    Bellicoso, C. D., Jenelten, F., Gehring, C. & Hutter, M. Dynamic locomotion through online nonlinear motion optimization for quadrupedal robots.IEEE Robotics and Automation Letters3, 2261–2268 (2018)

    work page 2018

  7. [7]

    & Semini, C

    Fahmi, S., Mastalli, C., Focchi, M. & Semini, C. Passive whole-body control for quadruped robots: Experimental validation over challenging terrain.IEEE Robotics and Automation Letters4, 2553–2560 (2019)

    work page 2019

  8. [8]

    & Kim, S

    Kim, D., Di Carlo, J., Katz, B., Bledt, G. & Kim, S. Highly dynamic quadruped locomotion via whole-body impulse control and model predictive control (2019). Preprint at http://arxiv.org/abs/1909. 06586

    work page 2019

Show all 62 references
  1. [9]

    Sutton, R. S. Reinforcement learning: An introduction.A Bradford Book(2018)

    2018

  2. [10]

    Tang, C.et al.Deep reinforcement learning for robotics: A survey of real-world successes.Annual Review of Control, Robotics, and Autonomous Systems8(2024)

    2024

  3. [11]

    M., Elyan, E

    Hussein, A., Gaber, M. M., Elyan, E. & Jayne, C. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR)50, 1–35 (2017)

    2017

  4. [12]

    & Hurst, J.Blind bipedal stair traversal via sim-to-real reinforcement learning(2021)

    Siekmann, J., Green, K., Warila, J., Fern, A. & Hurst, J.Blind bipedal stair traversal via sim-to-real reinforcement learning(2021). 15

    2021

  5. [13]

    Preprint at https://arxiv.org/ abs/2409.17731

    Vogel, D.et al.Robust ladder climbing with a quadrupedal robot (2024). Preprint at https://arxiv.org/ abs/2409.17731

    2024 arXiv

  6. [14]

    Han, L.et al.Lifelike agility and play in quadrupedal robots using reinforcement learning and generative pre-trained models.Nature Machine Intelligence6, 787–798 (2024)

    2024

  7. [15]

    & Kim, S

    Jang, Y., Seol, W., Lee, K., Kim, K.-S. & Kim, S. Development of quadruped robot for inspection of underground pipelines in nuclear power plants.Electronics Letters58, 234–236 (2022)

    2022

  8. [16]

    Deep reinforcement learning for multi-contact motion planning of hexapod robots, 2381–2388 (IJCAI, 2021)

    Fu, H.et al. Deep reinforcement learning for multi-contact motion planning of hexapod robots, 2381–2388 (IJCAI, 2021)

    2021

  9. [17]

    Wang, X.et al.Hierarchical free gait motion planning for hexapod robots using deep reinforcement learning.IEEE Transactions on Industrial Informatics19, 10901–10912 (2023)

    2023

  10. [18]

    Choi, S.et al.Learning quadrupedal locomotion on deformable terrain.Science Robotics8, eade2256 (2023)

    2023

  11. [19]

    D., Miki, T

    Fankhauser, P., Bjelonic, M., Bellicoso, C. D., Miki, T. & Hutter, M.Robust rough-terrain locomotion with a quadrupedal robot, 5761–5768 (IEEE, 2018)

    2018

  12. [20]

    & Hutter, M

    Tsounis, V., Alge, M., Lee, J., Farshidian, F. & Hutter, M. Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning.IEEE Robotics and Automation Letters5, 3699–3706 (2020)

    2020

  13. [21]

    & Pathak, D.Extreme parkour with legged robots, 11443–11450 (IEEE, 2024)

    Cheng, X., Shi, K., Agarwal, A. & Pathak, D.Extreme parkour with legged robots, 11443–11450 (IEEE, 2024)

    2024

  14. [22]

    & Hutter, M.Learning to walk in minutes using massively parallel deep reinforcement learning, 91–100 (PMLR, 2022)

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

    2022

  15. [23]

    & Hutter, M

    Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V. & Hutter, M. Learning quadrupedal locomotion over challenging terrain.Science Robotics5, eabc5986 (2020)

    2020

  16. [24]

    & Yang, W

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

    2022

  17. [25]

    & Stone, P.Behavioral cloning from observation, 4950–4957 (2018)

    Torabi, F., Warnell, G. & Stone, P.Behavioral cloning from observation, 4950–4957 (2018)

    2018

  18. [26]

    Peng, X. B.et al. Learning agile robotic locomotion skills by imitating animals(2020)

    2020

  19. [27]

    Imitating human behaviour with diffusion models(2023)

    Pearce, T.et al. Imitating human behaviour with diffusion models(2023)

    2023

  20. [28]

    & Levine, S

    Finn, C., Christiano, P., Abbeel, P. & Levine, S. A connection between generative adversarial networks, inverse reinforcement learning, and energy-based models (2016). Preprint at https://arxiv.org/abs/1611. 03852

    2016

  21. [29]

    & Levine, S.Learning robust rewards with adversarial inverse reinforcement learning (2018)

    Fu, J., Luo, K. & Levine, S.Learning robust rewards with adversarial inverse reinforcement learning (2018)

    2018

  22. [30]

    & Ermon, S

    Ho, J. & Ermon, S. Generative adversarial imitation learning.Advances in neural information processing systems29(2016)

    2016

  23. [31]

    Orsini, M.et al.What matters for adversarial imitation learning?Advances in Neural Information Processing Systems34, 14656–14668 (2021)

    2021

  24. [32]

    & Luo, Z.-Q

    Xu, T., Li, Z., Yu, Y. & Luo, Z.-Q. Understanding adversarial imitation learning in small sample regime: A stage-coupled analysis (2022). Preprint at https://arxiv.org/abs/2208.01899

    2022 arXiv

  25. [33]

    B., Ma, Z., Abbeel, P., Levine, S

    Peng, X. B., Ma, Z., Abbeel, P., Levine, S. & Kanazawa, A. Amp: Adversarial motion priors for stylized physics-based character control.ACM Transactions on Graphics (ToG)40, 1–20 (2021). 16

    2021

  26. [34]

    & Ermon, S

    Li, Y., Song, J. & Ermon, S. Infogail: Interpretable imitation learning from visual demonstrations. Advances in neural information processing systems30(2017)

    2017

  27. [35]

    B., Guo, Y., Halper, L., Levine, S

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

    2022

  28. [36]

    & Wang, W.C·ase: Learning conditional adversarial skill embeddings for physics-based characters, 1–11 (2023)

    Dou, Z., Chen, X., Fan, Q., Komura, T. & Wang, W.C·ase: Learning conditional adversarial skill embeddings for physics-based characters, 1–11 (2023)

    2023

  29. [37]

    Zhao, W., Queralta, J. P. & Westerlund, T.Sim-to-real transfer in deep reinforcement learning for robotics: a survey, 737–744 (IEEE, 2020)

    2020

  30. [38]

    B., Andrychowicz, M., Zaremba, W

    Peng, X. B., Andrychowicz, M., Zaremba, W. & Abbeel, P.Sim-to-real transfer of robotic control with dynamics randomization, 3803–3810 (IEEE, 2018)

    2018

  31. [39]

    Domain randomization for transferring deep neural networks from simulation to the real world, 23–30 (IEEE, 2017)

    Tobin, J.et al. Domain randomization for transferring deep neural networks from simulation to the real world, 23–30 (IEEE, 2017)

    2017

  32. [40]

    M.et al.Learning dexterous in-hand manipulation.The International Journal of Robotics Research39, 3–20 (2020)

    Andrychowicz, O. M.et al.Learning dexterous in-hand manipulation.The International Journal of Robotics Research39, 3–20 (2020)

    2020

  33. [41]

    & Chen, C.Easi: Evolutionary adversarial simulator identification for sim-to-real transfer, Vol

    Dong, H., Fu, H., Xu, W., Zhou, Z. & Chen, C.Easi: Evolutionary adversarial simulator identification for sim-to-real transfer, Vol. 37, 6603–6624 (2024)

    2024

  34. [42]

    Asid: Active exploration for system identification in robotic manipulation(2024)

    Memmel, M.et al. Asid: Active exploration for system identification in robotic manipulation(2024)

    2024

  35. [43]

    & Hilliges, O.Guiding infogan with semi-supervision, 119–134 (Springer, 2017)

    Spurr, A., Aksan, E. & Hilliges, O.Guiding infogan with semi-supervision, 119–134 (Springer, 2017)

    2017

  36. [44]

    Ess-infogail: Semi-supervised imitation learning from imbalanced demonstrations, Vol

    Fu, H.et al. Ess-infogail: Semi-supervised imitation learning from imbalanced demonstrations, Vol. 36, 60048–60059 (2023)

    2023

  37. [45]

    & Gomes, R

    Krause, A., Perona, P. & Gomes, R. Discriminative clustering by regularized information maximization. Advances in neural information processing systems23(2010)

    2010

  38. [46]

    & Kr¨ ahenb¨ uhl, P.Learning by cheating, 66–75 (PMLR, 2020)

    Chen, D., Zhou, B., Koltun, V. & Kr¨ ahenb¨ uhl, P.Learning by cheating, 66–75 (PMLR, 2020)

    2020

  39. [47]

    & Yu, Y.Hybrid actor-critic reinforcement learning in parameterized action space, 2279–2285 (IJCAI,2019)

    Fan, Z., Su, R., Zhang, W. & Yu, Y.Hybrid actor-critic reinforcement learning in parameterized action space, 2279–2285 (IJCAI,2019)

    2019

  40. [48]

    Grill, J.-B.et al.Bootstrap your own latent-a new approach to self-supervised learning.Advances in neural information processing systems33, 21271–21284 (2020)

    2020

  41. [49]

    & Zhang, Z.Acgail: Imitation learning about multiple intentions with auxiliary classifier gans, 321–334 (Springer, 2018)

    Lin, J. & Zhang, Z.Acgail: Imitation learning about multiple intentions with auxiliary classifier gans, 321–334 (Springer, 2018)

    2018

  42. [50]

    & Gong, Y.Document clustering based on non-negative matrix factorization, 267–273 (2003)

    Xu, W., Liu, X. & Gong, Y.Document clustering based on non-negative matrix factorization, 267–273 (2003)

    2003

  43. [51]

    Preprint at https://doi.org/10.48550/arXiv.2108.10470

    Makoviychuk, V.et al.Isaac gym: High performance gpu-based physics simulation for robot learning (2021). Preprint at https://doi.org/10.48550/arXiv.2108.10470

    2021 doi arXiv

  44. [52]

    & Agakov, F

    Barber, D. & Agakov, F. The im algorithm: a variational approach to information maximization. Advances in neural information processing systems16, 201 (2004)

    2004

  45. [53]

    Least squares generative adversarial networks, 2794–2802 (2017)

    Mao, X.et al. Least squares generative adversarial networks, 2794–2802 (2017)

    2017

  46. [54]

    Preprint at https://arxiv.org/abs/1801.00690

    Tassa, Y.et al.Deepmind control suite (2018). Preprint at https://arxiv.org/abs/1801.00690

    2018 arXiv

  47. [55]

    K., Goswami, S

    Ghosal, A., Nandy, A., Das, A. K., Goswami, S. & Panday, M. A short review on different clustering techniques and their applications.Emerging Technology in Modelling and Graphics: Proceedings of IEM 17 Graph 201869–83 (2020)

    2020

  48. [56]

    & Bengio, Y

    Grandvalet, Y. & Bengio, Y. Semi-supervised learning by entropy minimization.Advances in neural information processing systems17(2004)

    2004

  49. [57]

    & Saito, J

    Zhang, H., Starke, S., Komura, T. & Saito, J. Mode-adaptive neural networks for quadruped motion control.ACM Transactions on Graphics (TOG)37, 1–11 (2018)

    2018

  50. [58]

    & Klimov, O

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A. & Klimov, O. Proximal policy optimization algorithms (2017). Preprint at https://doi.org/10.48550/arXiv.1707.06347

    2017 doi arXiv

  51. [59]

    & Bengio, Y

    Chung, J., Gulcehre, C., Cho, K. & Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling (2014). Preprint at https://doi.org/10.48550/arXiv.1412.3555

    2014 doi arXiv

  52. [60]

    Paszke, A.et al.Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems32(2019)

    2019

  53. [61]

    Intel realsense.Intelhttps://www.intelrealsense.com/ (2024)

    2024

  54. [62]

    R.et al.Grad-cam: visual explanations from deep networks via gradient-based localization

    Selvaraju, R. R.et al.Grad-cam: visual explanations from deep networks via gradient-based localization. International journal of computer vision128, 336–359 (2020). 18 Supplementary Materials for Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robot...

    2020

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.