Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Learning Impact-Rich Rotational Maneuvers via Centroidal Velocity Rewards and Sim-to-Real Techniques: A One-Leg Hopper Flip Case Study

T0 review · 3 major / 4 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 reward built on the whole body's centroidal angular velocity, not base-link spin or total angular momentum, is what makes a one-leg hopper front-flip.

desk verdict Genuinely useful CAV reward and a real hardware flip, but the transmission-load-regularization evidence rests on an unvalidated torque estimator. read the letter →

arxiv 2505.12222 v3 pith:DZSDB4X5 submitted 2025-05-18 cs.RO

Dongyun Kang , Gijeong Kim , JongHun Choe , Hajun Kim , Hae-Won Park This is my paper · ORCID
classification cs.RO
keywords reinforcementlearningsim-to-realtransfercentroidalangularvelocityfrontflipone-leghoppermotoroperatingregiontransmissionloadregularizationleggedrobots
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 claims that to make a legged robot perform a true full-body rotation, the learning objective must reward the system's centroidal angular velocity—the overall rotation rate of the whole body about its center of mass—rather than the base link's angular velocity or the total angular momentum. On a 12.45 kg one-leg hopper, the authors show that maximizing base angular velocity produces only thigh-calf flailing with no takeoff, while maximizing angular momentum generates takeoff but too little rotation; only the centroidal velocity reward yields a complete front flip. The paper also argues that two actuator-aware sim-to-real techniques are needed for hardware transfer: modeling the motor operating region in the torque–speed plane to keep torque commands feasible, and regularizing transmission load so landing impacts do not fracture the ankle's sun gear. With these ingredients, the authors report the first hardware realization of a full front flip on a one-leg hopper, over eight successful trials.

What carries the argument

The load-bearing object is the centroidal angular velocity reward, defined as $r_{\mathrm{CAV}} = \max(\min(\alpha^T \omega_{\mathrm{com}}, 10), -0.1)$ during the aerial phase, where $\omega_{\mathrm{com}}$ is the centroidal angular velocity obtained from the centroidal momentum relation $h_G = I_G v_G$ and $\omega_{\mathrm{com}} = I_{\mathrm{com}}^{-1} L_{\mathrm{com}}$. This quantity links momentum to posture-dependent inertia, so rewarding it inherently encourages both momentum generation and mid-air inertia reduction. The supporting machinery is made of Motor Operating Region modeling, which clips commanded torques to a trapezoidal torque–speed envelope bounded by the voltage-limit slope and the current limit, and transmission load regularization, which penalizes contact-derived joint loads through a relaxed log barrier and probabilistic episode termination when loads cross a critical threshold.

What would settle it

Instrument the physical ankle joint with a strain gauge or torque sensor and compare peak landing loads between regularized and unregularized policies over many trials. The paper's claim predicts a peak external torque drop from roughly 21 N·m to 17 N·m and no sun gear fracture across at least eight flips; if measured peak loads are statistically unchanged or the regularized hardware still fractures, the transmission-load benefit fails.

Watch

Extended reading notes

Core claim

The central discovery is that the choice of what quantity is rewarded determines whether whole-body rotation actually emerges. For rotational maneuvers, the centroidal angular velocity $w_{\mathrm{com}} = I_{\mathrm{com}}^{-1} L_{\mathrm{com}}$ couples angular momentum with the posture-dependent composite inertia, so maximizing it rewards both the generation of momentum and the reduction of inertia through configuration change. The paper shows this leads to a policy that extends the leg to build momentum during takeoff, then tucks the knee mid-air to raise spin rate from 7.4 rad/s to 10.6 rad/s and complete a full flip. By contrast, base angular velocity can be gamed by internal joint motion, and momentum-only rewards leave the leg extended and the rotation too slow. Combined with Motor Operating Region clipping and transmission load regularization, this reward is what the authors credit for the successful hardware transfer and repeated flip execution on the one-leg hopper.

Load-bearing premise

The load-reduction result depends on the simplified inertia model used to estimate ankle external torque on hardware, $I_{\mathrm{eff}} = I_{\mathrm{rotor}} + I_{\mathrm{foot}}/2$; that estimator is validated only in simulation, so if it is inaccurate on the real closed-loop mechanism, the claimed torque drop and durability gain may not hold.

Editorial extensions

If this is right

  • A CAV-based reward, with MOR and load regularization, produces a complete front flip on real hopper hardware, whereas base-angular-velocity policies never leave the ground and angular-momentum policies undershoot the rotation.
  • Policies trained with MOR constraints keep torque commands inside the feasible actuator envelope, while policies trained with only box-shaped torque limits issue unattainable commands and fail when evaluated under MOR.
  • Transmission load regularization cuts the estimated peak ankle external torque from about 21 N·m to 17 N·m and lets the robot complete eight consecutive hardware flips, where the unregularized policy fractured the sun gear on the second trial.
  • The reward naturally produces a mid-air knee tuck that reduces composite inertia and raises centroidal angular velocity by 43%, from 7.4 rad/s at takeoff to 10.6 rad/s in flight.
  • The same reward formulation, with the flip axis as a parameter, also learns yaw spins and barrel rolls on the hopper and a backflip on a quadruped in simulation, suggesting the recipe extends beyond the demonstrated maneuver.

Reading between the lines

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

  • The BAV-versus-CAV comparison suggests a general diagnostic for acrobatic learning: if a policy spends energy on internal joint motion without global rotation, the reward should be moved from link rates to centroidal rates; the paper only demonstrates this for flips and spins, but the diagnosis is transferable.
  • Because CAV rewards both momentum and inertia reduction, a momentum-only policy could plausibly be repaired by adding an explicit inertia-reduction bonus; this is a testable variant the paper does not run.
  • MOR clipping acts as a policy regularization that may benefit other high-torque, high-speed behaviors such as sprinting and jumping, where the torque–speed tradeoff is also decisive; the paper leaves that application unexplored.
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

3 major / 4 minor

Summary. The paper proposes a centroidal angular velocity (CAV) reward for learning whole-body rotational maneuvers, combined with Motor Operating Region (MOR) modeling and transmission load regularization for sim-to-real transfer. As a case study, the authors train a one-leg hopper to perform a front flip, evaluate the reward design in simulation against base angular velocity (BAV) and centroidal angular momentum (CAM) rewards, compare MOR-constrained versus unconstrained policies, and present hardware experiments including repeated flips. They report the first hardware front flip on a one-leg hopper and attribute improved hardware durability to transmission load regularization.

Significance. If validated, the CAV reward is an elegant and transferable alternative to link-level rewards: it directly rewards the quantity that must be maximized for rotation while implicitly encouraging inertia reduction. The hardware front flip on a minimal single-foot platform is a notable advance, and the MOR analysis clearly shows the importance of actuator-aware constraints. The paper's strengths include clean ablations (BAV/CAM/CAV), simulation-to-hardware consistency for the MOR-evaluated policy, and reproducible details (reward tables, barrier formulation, domain randomization ranges). The main weaknesses concern the quantitative hardware evidence for transmission load regularization, which relies on an approximate external-torque estimator and a single baseline trial.

major comments (3)
  1. [§5.3, Appendix A.5] The hardware reduction in peak external ankle torque (from about 21 Nm to 17 Nm, Fig. 7c–d) is estimated using τ_ext = τ_input − I_eff·dω_rotor with I_eff = I_rotor + I_foot/2, an approximation for the closed-loop ankle mechanism that is validated only against simulation. The estimator's RMSE in the initial impact phase is 1.783 Nm, which is comparable to the claimed 4 Nm reduction, and the two policies land differently (flat-foot vs rolling), so a small bias in I_eff could alter the two estimates in opposite directions. Please validate the estimator on hardware (e.g., by applying known external torques or using an instrumented ankle) or provide a sensitivity analysis over I_eff to bound the uncertainty of the reported peak-torque reduction.
  2. [§5.3, Fig. 8] The durability conclusion—that transmission load regularization prevents sun gear fracture—rests on a single successful baseline trial before the fracture (n=1) versus eight trials for the regularized policy. This is anecdotal as reported. If additional baseline trials are not feasible, the paper should either present them or explicitly label the hardware durability evidence as a case observation rather than a demonstrated effect, and give the simulation results (Fig. 7a–b) the primary evidentiary weight.
  3. [§5.1, Fig. 4d–f] The mechanistic claim that CAM rewards fail because they do not incentivize inertia reduction is confounded: the CAM policy also produces substantially lower centroidal angular momentum (4.3 N·s) than the CAV policy (6.0 N·s). Thus the difference in centroidal angular velocity (4.8 vs 10.6 rad/s) could be partly due to lower momentum generation, not solely to the lack of inertia modulation. Please provide a matched-momentum comparison or a quantitative decomposition (e.g., reporting L and I at peak ω) to support the stated mechanism.
minor comments (4)
  1. [Section 2 heading] The heading 'Centroidal Momemtum' should be 'Centroidal Momentum'.
  2. [References] In Section 2, reference [11] is cited as 'Zhou et al.' but the reference list assigns [11] to Chignoli et al.; the intended 'Zhou et al.' appears to be [23]. Please correct the citation.
  3. [Title and Abstract] The title and abstract contain 'V elocity' with an extra space; fix the typography.
  4. [Figure 6 caption] The caption 'Motor-side torque (without gear reduction) is shown' is ambiguous because the figure plots three quantities; please clarify which curve corresponds to motor-side torque.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the centroidal angular velocity reward is a training input, the hardware flip is an external benchmark, and the cited MOR and barrier tools are stated explicitly rather than used as load-bearing self-citation.

full rationale

Walking the claimed derivation chain: the centroidal angular velocity reward is defined from standard centroidal dynamics (w_com = I_com^{-1} L_com, following Orin and Goswami), and the aerial-phase reward is simply a clipped alignment of alpha^T w_com. This reward is an input to PPO training, not a quantity fitted to the observed flip outcome; the hardware deployment and the BAV/CAM/CAV comparisons are external benchmarks of the trained policies. No equation in the paper defines the prediction in terms of the reward, and the comparisons use distinct reward formulations rather than the proposed one. The MOR and transmission-load-regularization techniques cite prior work from the same group, but the paper explicitly states the MOR envelope (voltage-limit slope and current limit) and the relaxed log barrier formula in Appendix A.2, so these citations are attributions of tools rather than load-bearing self-citation. The external torque estimator in Appendix A.5 is an approximation (I_eff = I_rotor + I_foot/2) and is validated only in simulation, with RMSE 1.783 Nm during the initial impact phase; this is a legitimate correctness and robustness concern about the claimed 21-to-17 Nm reduction, but it is not circularity, because the estimator is not constructed to force that reduction. The Limitations section narrows the generality claims but does not reveal any step where a fitted parameter is renamed as a prediction or where an input defines the output. Overall, the central contributions are self-contained against external benchmarks, and no specific reduction of a claimed result to its inputs is exhibited.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard centroidal dynamics and a set of hand-chosen reward and regularization terms. No new physical entity is introduced. The MOR envelope and transmission load estimates are derived from hardware measurements and a simplified inertia model, respectively; the external torque estimator is the least independently verified component and is treated here as a domain assumption.

free parameters (6)
  • Aerial phase boundaries = 0.5 to 1.05 seconds
    The takeoff, aerial, and landing phase times in the CAV reward are hand-set to the hopper's observed flip timing.
  • CAV reward cap and floor = max 10, min -0.1
    The aerial-phase reward saturates at 10 rad/s and has a small negative floor; these values are chosen by hand to normalize the reward scale.
  • Landing momentum penalty scale = -0.5 * min(||L_com||, 2.5)
    The landing-phase penalty coefficients are hand-tuned to encourage stopping rotation without destabilizing early training.
  • Reward weights in Table 1 = See Table 1 values
    Weights for linear velocity, torque, joint position, velocity, acceleration, slip, action smoothness, contact smoothness, and contact impulse are hand-tuned regularizers.
  • Transmission load threshold and termination probability = Critical threshold, 50% termination
    The episode termination rule for excessive transmission load uses a hand-chosen critical threshold and a 50% probability.
  • Domain randomization ranges = Mass 0.85-1.15, COM shift +/-3 cm, friction 0.5-0.9, etc.
    The randomization ranges in Table 2 are chosen manually to approximate hardware variability.
assumptions (5)
  • standard math Centroidal momentum relation h_G = I_G v_G and conservation of centroidal angular momentum during flight
    The reward is based on the standard centroidal dynamics formulation of Orin and Goswami, assumed valid for the rigid multibody model.
  • domain assumption RaiSim pin constraints model the closed-loop ankle mechanism accurately
    Section 3 states the closed-loop ankle is simulated in RaiSim with pin constraints; the sim-to-real transfer relies on this model being faithful enough for the learned policy.
  • domain assumption The MOR envelope with voltage limit slope and current limit captures real actuator torque-speed limits
    Section 4.2 assumes the trapezoidal MOR computed from motor operating data represents the true feasible torque region of the hardware.
  • domain assumption The external torque estimator tau_ext = tau_input - I_eff * domega_rotor, with I_eff = I_rotor + I_foot/2, is valid on hardware
    Appendix A.5 uses this approximation to estimate landing loads on the sun gear; it is validated in simulation but not directly on hardware.
  • domain assumption PD gain tracking with fixed gains is an adequate low-level control layer for the learned policy
    The policy outputs joint position targets and a PD controller generates torques; the success of the method depends on the PD layer tracking commands closely enough in both simulation and hardware.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Impact-Rich Rotational Maneuvers via Centroidal Velocity Rewards and Sim-to-Real Techniques: A One-Leg Hopper Flip Case Study." pith.science (2026). https://pith.science/paper/DZSDB4X5

@misc{pith2026250512222,
  author       = {Pith},
  title        = {Pith review of: Learning Impact-Rich Rotational Maneuvers via Centroidal Velocity Rewards and Sim-to-Real Techniques: A One-Leg Hopper Flip Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZSDB4X5}},
  note         = {Machine review of arXiv:2505.12222}
}
read the original abstract

Dynamic rotational maneuvers, such as front flips, inherently involve large angular momentum generation and intense impact forces, presenting major challenges for reinforcement learning and sim-to-real transfer. In this work, we propose a general framework for learning and deploying impact-rich, rotation-intensive behaviors through centroidal velocity-based rewards and actuator-aware sim-to-real techniques. We identify that conventional link-level reward formulations fail to induce true whole-body rotation and introduce a centroidal angular velocity reward that accurately captures system-wide rotational dynamics. To bridge the sim-to-real gap under extreme conditions, we model motor operating regions (MOR) and apply transmission load regularization to ensure realistic torque commands and mechanical robustness. Using the one-leg hopper front flip as a representative case study, we demonstrate the first successful hardware realization of a full front flip. Our results highlight that incorporating centroidal dynamics and actuator constraints is critical for reliably executing highly dynamic motions. A supplementary video is available at: https://youtu.be/atMAVI4s1RY

Figures

Figures reproduced from arXiv: 2505.12222 by the authors.

Figure 1
Figure 1. Snapshots of the first successful deployment of the learned front flip on the real one-leg [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. One￾leg hopper used in this study. Our target platform is a compact, custom-designed 3-DOF one-leg hopper sys￾tem inspired by human lower-limb kinematics. It consists of a single knee pitch joint and a two-axis ankle (pitch and roll), terminating in a flat foot. The platform is capable of executing forward, backward, and lateral command-tracking loco￾motion, and can recover from external pushes with no additional su… view at source ↗
Figure 3
Figure 3. Motor Operating Region (MOR) at the knee actuator. Sim￾ulation data were collected within ±0.25 seconds during the flip mo￾tion around take-off, with take-off and knee fold events also noted in the figure. Red regions denote ar￾eas beyond the MOR. Motor Operating Region. Dynamic, impact-rich maneuvers demand torque performance at the very edge of actuator ca￾pabilities. In particular, front flips on a one leg hopper… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Comparison of policies trained with different aerial-phase rewards (0.5–1.05 s) for front [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of simulation and hardware results for policies with and without MOR con [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Experimental torque pro￾files at the ankle left motor during landing. Motor-side torque (without gear reduction) is shown. External Torque at Landing. In hardware trials without transmission load regularization, the robot completed one front flip, but the ankle actuato…
Figure 7
Figure 7. Figure 7: External torques on the ankle actuator’s sun gear during landing, for policies trained [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Pitch rotation across repeated hardware flip trials. The baseline (with￾out load regularization) failed on the second trial due to sun gear fracture (in￾set photo); the regularized policy com￾pleted eight trials without failure. The effect of transmission load regulari…
Figure 9
Figure 9. Figure 9: Comparison of estimated and simulated external torque on the ankle actuator across six [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Additional maneuvers learned by the one-leg hopper using the same reward framework. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: A backflip trained on the Unitree Go1 quadruped. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 28 canonical work pages

  1. [1]

    Y . Jin, X. Liu, Y . Shao, H. Wang, and W. Yang. High-speed quadrupedal locomotion by imitation-relaxation reinforcement learning. Nature Machine Intelligence , 4(12):1198–1208, 2022

    work page 2022

  2. [2]

    Shin, T.-G

    Y .-H. Shin, T.-G. Song, G. Ji, and H.-W. Park. Reinforcement learning for high-speed quadrupedal locomotion with motor operating region constraints: Mitigating motor model discrepancies through torque clipping in realistic motor operating region. IEEE Robotics & Automation Magazine, 2024

    work page 2024

  3. [3]

    Hoeller, N

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter. Anymal parkour: Learning agile navigation for quadrupedal robots. Science Robotics, 9(88):eadi7566, 2024

    2024

  4. [4]

    Dynamics

    B. Dynamics. Atlas — partners in parkour. URL https://www.youtube.com/watch?v= tF4DML7FIWk&ab_channel=BostonDynamics

    work page

  5. [5]

    Robotics

    U. Robotics. World’s first side-flipping humanoid robot: Unitree g1. URL https://www. youtube.com/watch?v=29xLWhqME2Q&ab_channel=UnitreeRobotics

    work page

  6. [6]

    Dynamics

    B. Dynamics. Walk, run, crawl, rl fun — boston dynamics — atlas. URL https://www. youtube.com/watch?v=I44_zbEwz_w&ab_channel=BostonDynamics

    work page

  7. [7]

    D. Kim, H. Kwon, J. Kim, G. Lee, and S. Oh. Stage-wise reward shaping for acro- batic robots: A constrained multi-objective reinforcement learning approach. arXiv preprint arXiv:2409.15755, 2024

    arXiv 2024

  8. [8]

    Li and P

    H. Li and P. M. Wensing. Cafe-mpc: A cascaded-fidelity model predictive control framework with tuning-free whole-body control. IEEE Transactions on Robotics, 2024

    work page 2024

Show all 37 references
  1. [9]

    S. Ha, J. Lee, M. van de Panne, Z. Xie, W. Yu, and M. Khadiv. Learning-based legged locomo- tion: State of the art and future perspectives. The International Journal of Robotics Research , page 02783649241312698, 2024

    2024

  2. [10]

    H. Dai, A. Valenzuela, and R. Tedrake. Whole-body motion planning with centroidal dynamics and full kinematics. In 2014 IEEE-RAS International Conference on Humanoid Robots , pages 295–302. IEEE, 2014. 9

    2014

  3. [11]

    Chignoli, D

    M. Chignoli, D. Kim, E. Stanger-Jones, and S. Kim. The mit humanoid robot: Design, motion planning, and control for acrobatic behaviors. In 2020 IEEE-RAS 20th International Confer- ence on Humanoid Robots (Humanoids) , pages 1–8. IEEE, 2021

    2020

  4. [12]

    Ostyn, B

    F. Ostyn, B. Vanderborght, and G. Crevecoeur. Improving the collision tolerance of high- speed industrial robots via impact-aware path planning and series clutched actuation. IEEE Transactions on Robotics, 2024

    2024

  5. [13]

    H.-W. Park, P. M. Wensing, and S. Kim. High-speed bounding with the mit cheetah 2: Con- trol design and experiments. The International Journal of Robotics Research , 36(2):167–192, 2017

    2017

  6. [14]

    G. Kim, D. Kang, J.-H. Kim, S. Hong, and H.-W. Park. Contact-implicit model predictive con- trol: Controlling diverse quadruped motions without pre-planned contact modes or trajectories. The International Journal of Robotics Research , 44(3):486–510, 2025

    2025

  7. [15]

    Caluwaerts, A

    K. Caluwaerts, A. Iscen, J. C. Kew, W. Yu, T. Zhang, D. Freeman, K.-H. Lee, L. Lee, S. Sal- iceti, V . Zhuang, et al. Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654, 2023

    2023 arXiv

  8. [16]

    Zhuang, Z

    Z. Zhuang, Z. Fu, J. Wang, C. G. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao. Robot parkour learning. In Conference on Robot Learning, pages 73–92. PMLR, 2023

    2023

  9. [17]

    Dynamics

    B. Dynamics. What’s new, atlas? URL https://www.youtube.com/watch?v= fRj34o4hN4I&ab_channel=BostonDynamics

  10. [18]

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

    2023

  11. [19]

    Fuchioka, Z

    Y . Fuchioka, Z. Xie, and M. Van de Panne. Opt-mimic: Imitation of optimized trajectories for dynamic quadruped behaviors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5092–5098. IEEE, 2023

    2023

  12. [20]

    M. I. of Technology (MIT). Backflipping mit mini cheetah. URL https://www.youtube. com/watch?v=xNeZWP5Mx9s&ab_channel=MassachusettsInstituteofTechnology% 28MIT%29

  13. [21]

    Lee and A

    S.-H. Lee and A. Goswami. Reaction mass pendulum (rmp): An explicit model for centroidal angular momentum of humanoid robots. In Proceedings 2007 IEEE international conference on robotics and automation, pages 4667–4672. IEEE, 2007

    2007

  14. [22]

    D. E. Orin and A. Goswami. Centroidal momentum matrix of a humanoid robot: Structure and properties. In 2008 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages 653–659. IEEE, 2008

    2008

  15. [23]

    Z. Zhou, B. Wingo, N. Boyd, S. Hutchinson, and Y . Zhao. Momentum-aware trajectory opti- mization and control for agile quadrupedal locomotion. IEEE Robotics and Automation Let- ters, 7(3):7755–7762, 2022

    2022

  16. [24]

    P. M. Wensing, M. Posa, Y . Hu, A. Escande, N. Mansard, and A. Del Prete. Optimization-based control for dynamic legged robots. IEEE Transactions on Robotics, 40:43–63, 2023

    2023

  17. [25]

    Vezzi, J

    F. Vezzi, J. Ding, A. Raffin, J. Kober, and C. Della Santina. Two-stage learning of highly dynamic motions with rigid and articulated soft quadrupeds. In 2024 IEEE International Con- ference on Robotics and Automation (ICRA) , pages 9720–9726. IEEE, 2024

    2024

  18. [26]

    Ferigo, R

    D. Ferigo, R. Camoriano, P. M. Viceconte, D. Calandriello, S. Traversaro, L. Rosasco, and D. Pucci. On the emergence of whole-body strategies from humanoid robot push-recovery learning. IEEE Robotics and Automation Letters , 6(4):8561–8568, 2021. 10

    2021

  19. [27]

    Zhang, X

    X. Zhang, X. Wang, L. Zhang, G. Guo, X. Shen, and W. Zhang. Achieving stable high- speed locomotion for humanoid robots with deep reinforcement learning. arXiv preprint arXiv:2409.16611, 2024

    2024 arXiv

  20. [28]

    W. Xie, C. Bai, J. Shi, J. Yang, Y . Ge, W. Zhang, and X. Li. Humanoid whole-body loco- motion on narrow terrain via dynamic balance and reinforcement learning. arXiv preprint arXiv:2502.17219, 2025

    2025 arXiv

  21. [29]

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, and V . Vanhoucke. Sim- to-real: Learning agile locomotion for quadruped robots. arXiv preprint arXiv:1804.10332 , 2018

    2018 arXiv

  22. [30]

    Bin Peng, E

    X. Bin Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine. Learning agile robotic locomotion skills by imitating animals. Robotics: Science and Systems (RSS), Virtual Event/Corvalis, July, pages 12–16, 2020

    2020

  23. [31]

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

    2020

  24. [32]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning robust percep- tive locomotion for quadrupedal robots in the wild. Science robotics, 7(62):eabk2822, 2022

    2022

  25. [33]

    Kim, Y .-H

    G. Kim, Y .-H. Lee, and H.-W. Park. A learning framework for diverse legged robot locomotion using barrier-based style rewards. arXiv preprint arXiv:2409.15780, 2024

    2024 arXiv

  26. [34]

    J. Chen, J. Frey, R. Zhou, T. Miki, G. Martius, and M. Hutter. Identifying terrain physical parameters from vision-towards physical-parameter-aware locomotion and navigation. IEEE Robotics and Automation Letters, 2024

    2024

  27. [35]

    H. Kim, D. Kang, M.-G. Kim, G. Kim, and H.-W. Park. Online friction coefficient identifi- cation for legged robots on slippery terrain using smoothed contact gradients. IEEE Robotics and Automation Letters, 2025

    2025

  28. [36]

    Hwangbo, J

    J. Hwangbo, J. Lee, and M. Hutter. Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters , 3(2):895–902, 2018

    2018

  29. [37]

    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 , 7(2):4630–4637, 2022. 11 A Appendix A.1 Reward Formulations Table 1: Reward terms with expressi...

    2022

Pith tools

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