Pith. sign in

REVIEW 3 major objections 7 minor 28 references

Using coordinated foreleg manipulation and hind-leg support, a simulation-trained hierarchical controller enables a hexapod to push objects of unknown size, weight, and friction to specified poses on real hardware.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 14:38 UTC pith:C4RLKUOR

load-bearing objection A credible hexapod pushing system with strong simulation evidence but anecdotal real-world validation; the central claim outruns the hardware data. the 3 major comments →

arxiv 2509.23651 v3 pith:C4RLKUOR submitted 2025-09-28 cs.RO

HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation by a Hexapod Robot

Xinrong Yang , Peizhuo Li , Hongyi Li , Yifeng Peng , Arhaan Jain , Junkai Lu , Linnan Chang , Yuhong Cao
show 3 more authors
Yifeng Zhang Ge Sun Guillaume Sartoretti
This is my paper
classification cs.RO
keywords hexapod robotwhole-body loco-manipulationnon-prehensile manipulationhierarchical reinforcement learningsim-to-real transferobject pushingdomain randomizationmulti-leg coordination
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 paper claims that hexapod robots can perform whole-body loco-manipulation of objects comparable to their own body size and weight without a dedicated arm or gripper, using only coordinated leg movements. It introduces HeLoM, which splits the task into a high-level planner that outputs pushing commands (body velocity and foreleg foot targets) and a low-level controller that maintains balance and executes those commands. Both are trained with reinforcement learning in simulation, with heavy domain randomization over object properties and joint gains, and are deployed zero-shot on a physical hexapod. Real-world experiments show the robot pushing four different boxes, including a heavy one with a shifted center of mass and an uneven surface never seen in training, to specified poses, even under human-imposed disturbances. If this holds, it suggests a practical path to multi-legged manipulation without extra hardware.

Core claim

The central claim is that whole-body, non-prehensile pushing by a hexapod can be decomposed into two learned policies that work together: a Planner that decides where to place the forelegs and how to move the body, and a Controller that tracks those decisions while keeping the robot upright. The Planner is rewarded for bringing the object to the goal, but only after both forelegs are in contact, and it also receives penalties for object tilt, foot slip, and erratic actions. The Controller is trained first with smoothness regularization and domain randomization, then frozen while the Planner is trained on top. The paper reports a 95.6% success rate over 1000 simulated trials and successful re

What carries the argument

The load-bearing mechanism is the hierarchical decomposition into Planner and Controller networks, both trained with reinforcement learning. The Planner runs at 50 Hz and outputs desired body velocities and target foreleg foot positions; the Controller maps proprioceptive observations and latent estimates to 18 joint position targets. Three design choices carry the argument: a reward that gives a bonus for simultaneous contact of both forelegs, encouraging coordinated bi-foot pushing; a term that penalizes object tilt around the x/y axes to keep the pushing line near the center of mass and prevent tipping; and a curriculum over target lateral displacement and orientation that starts easy and

Load-bearing premise

The central claim rests on the assumption that randomized physics parameters and PD gain scaling in simulation adequately model the real hexapod's Series Elastic Actuator nonlinearities, so a policy trained entirely in simulation with no fine-tuning will behave on the real robot as it does in simulation.

What would settle it

Deploy the exact HeLoM policy on a real box with mass 10 kg and surface friction around 0.5 (well inside the randomized ranges) and measure success rate over 20 independent trials; if the real-world success is below 50% while simulation reports 95.6% for the same distribution, the zero-shot transfer claim collapses.

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

If this is right

  • Hexapod robots can perform useful manipulation of large, heavy, irregular objects without a dedicated arm, using only coordinated leg forces.
  • Zero-shot sim-to-real transfer is feasible for contact-rich pushing on platforms with compliant actuators, provided the domain randomization is broad enough.
  • The hierarchical planner/controller split may generalize to other multi-legged platforms and to non-prehensile tasks beyond pushing, such as rolling or dragging.
  • A command curriculum is necessary for learning long-range pushing with large orientation changes; without it, simulated success drops from 95.6% to 78.7%.
  • The robot can adapt online to external disturbances and object geometry it has never seen, suggesting the learned policy is not merely memorizing a single pushing strategy.

Where Pith is reading between the lines

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

  • If the framework's success stems from the hexapod's redundant legs, the same hierarchy may transfer to other legged robots, but with fewer legs the stability margin would be thinner; a direct comparison would test that hypothesis.
  • The paper's reported failures on lightweight, low-friction objects suggest the policy lacks fine force control at the low end of the distribution; adding force feedback or a different terminal reward is a testable fix.
  • The reliance on an external motion capture system means the perception problem is deferred; a version with onboard cameras or LiDAR would need to tolerate pose-estimation noise, which is an open extension not addressed here.
  • A conservative reading of 'unknown physical properties' is that they vary within the randomized ranges; the paper does not claim generalization to, say, 30 kg boxes or very slippery surfaces.

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 / 7 minor

Summary. The paper proposes HeLoM, a hierarchical reinforcement-learning framework for whole-body loco-manipulation by a hexapod robot. A high-level Planner outputs commanded base velocities and foreleg foot positions; a low-level Controller, based on DreamWaQ, tracks these commands while maintaining locomotion stability. Both levels are trained in Isaac Gym with domain randomization and a command-based curriculum, then deployed zero-shot on the HEBI Daisy hexapod. Simulation evaluation over 1000 randomized trials reports a 95.6% success rate with an average completion time of 12.1 s, alongside ablations showing the importance of the curriculum and key reward terms. Real-world experiments qualitatively demonstrate pushing of four boxes with varying sizes, masses, materials, and center-of-mass locations, as well as recovery from external disturbances. The central claim is that the framework can stably push objects of varying sizes and unknown physical properties to designated goal poses in the real world.

Significance. If the central claim is fully substantiated, HeLoM would be a meaningful advance in legged loco-manipulation: it is among the first hexapod-specific frameworks to use coordinated foreleg interaction and hind-leg propulsion for non-prehensile pushing, requiring no dedicated arm and no real-world fine-tuning. The paper has clear strengths: the hierarchical decomposition with a frozen low-level controller is principled; the command-based curriculum is a sensible response to sparse exploration; the domain randomization over physical parameters, PD gains, and object properties is well motivated; and the simulation study is strong in scope, with 1000 randomized trials and ablations isolating the curriculum and key rewards. The main weakness is that the paper's headline claim is about real-world capability, but the real-world evidence is purely qualitative. The simulation metrics cannot by themselves establish the sim-to-real transfer claims, especially given the paper's own statement that Daisy's series elastic actuators enlarge the sim-to-real gap.

major comments (3)
  1. [Section IV.B and Figs. 5-7] The central abstract claim — that HeLoM can 'stably push objects of varying sizes and unknown physical properties to designated goal poses in the real world' — is not quantitatively supported by the hardware section. Figs. 5-7 show individual qualitative sequences, but there is no trial count, no success rate, no final pose error, no completion time, and no comparison condition on hardware. Table VI reports a 95.6% success rate over 1000 simulation trials, but this cannot be extrapolated to the real robot. The authors should define the same success criteria used in simulation (position error below 0.05 m, orientation error below 5 degrees) and report success rate, error statistics, and trial counts for the real-world pushing task, including the disturbance-recovery scenario.
  2. [Section III.C.1 and Table IV] The paper states that Daisy's Series Elastic Actuators have nonlinearities and parameter uncertainties that 'further enlarge the sim-to-real gap,' and the proposed mitigation is domain randomization over PD gains and physical parameters. However, no quantitative evidence is provided that this mitigation suffices on hardware. Fig. 4 shows a qualitative velocity/foot-tracking comparison with and without the smoothness loss, but it does not measure task-level pushing performance. The zero-shot deployment claim depends on the effectiveness of this transfer mechanism, so the hardware evaluation should include quantitative tracking errors and task success rates, ideally with a comparison against a non-adapted baseline, to demonstrate that the domain randomization actually closes the gap.
  3. [Section IV.B.1-2 and Section V] The real-world experiments rely on an external optical motion capture system for object and robot poses, and the conclusion explicitly concedes that this 'limits its use to controlled environments.' The abstract, however, presents the real-world result without this caveat. This is an evidence gap rather than an internal inconsistency, but it should be addressed: either the abstract should qualify the claim to 'controlled indoor environments with external pose feedback,' or the paper should provide additional real-world evidence showing the controller works when such external infrastructure is unavailable.
minor comments (7)
  1. [Title] The title reads 'in Hexapod Robot'; grammatically it should be 'for a Hexapod Robot' or 'with a Hexapod Robot.'
  2. [Section III.C.2] The command-based curriculum is described only qualitatively. For reproducibility, specify the number of curriculum levels, the exact thresholds for y-displacement and orientation, and the expansion schedule.
  3. [Section III.C.1, Eq. (2)] The smoothness-loss equation uses hyperparameters λ_π and λ_V but their values are not reported. Please provide the values used in training.
  4. [Table II] The notation `rdis,obj tar` appears with inconsistent subscripts, and the penalty term `P([a_min - a_t]^+ + [a_t - a_max]^+)` is not formally defined. Clarify whether this is an indicator or a hinge penalty and define the bracket notation.
  5. [Fig. 4] The velocity and foot-position tracking plots lack explicit axis labels, units, and quantitative error metrics. Adding these would make the comparison with and without smoothness loss more informative.
  6. [Section IV.B.2] The sentence 'e) is the object specifications and experimental test parameters' is a sentence fragment; rephrase. Also, consider stating explicitly whether the four boxes in Fig. 6 are the same as those used for the quantitative evaluation in Table VI, if applicable.
  7. [Section II.A] There is a typographical error: 'B’ezier' should be 'Bézier.'

Circularity Check

0 steps flagged

No significant circularity: the paper's claims rest on trained-policy evaluation and external baselines, not on self-referential derivation.

full rationale

HeLoM is an empirical RL system: the high-level Planner and low-level Controller are trained in Isaac Gym with PPO, and the paper's central claims are supported by held-out simulation success rates (95.6% over 1000 randomized trials, Table VI) and by qualitative real-world pushing demonstrations (Figs. 5–7). There is no analytic derivation chain in which an output is defined in terms of an input, and no fitted parameter is later renamed as a prediction. The simulation success criterion (position error < 0.05 m, orientation error < 5°) is the same task definition used in training, but that is standard, disclosed benchmarking rather than circularity: the RL policy is not constructed to satisfy that criterion by hand; it is optimized against a reward function and then evaluated under the same task metric, which is an empirical measurement. Domain randomization over friction, mass, COM, inertia, and PD gains (Tables IV–V) is a known sim-to-real mitigation, not a hidden way of encoding the real-world result. Self-citations appear in related work ([1], [4], [19]) and are not load-bearing for the central claim; the controller architecture is taken from the external DreamWaQ work [25], and the smoothness loss from external L2C2 [27]. The paper's own conclusion explicitly concedes that the current implementation requires an external motion capture system, which limits deployment to controlled environments; this is a limitation disclosure, not a circular step. The skeptic's concern—that real-world support is anecdotal without quantitative success rates or pose errors—is an evidence-strength issue, not a circularity issue. I find no step where a prediction reduces by construction to its inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central contribution is an engineered control system, not a derivation. The numbers the authors control are reward weights, command bounds, curriculum design, success thresholds, and domain randomization ranges; all are hand-chosen and are itemized above. The method relies on standard algorithms (PPO, Isaac Gym, DreamWaQ, L2C2 smoothness) as background assumptions rather than introducing new physical entities. No new particles, forces, or conserved quantities are postulated.

free parameters (5)
  • Planner reward weights = Table II; e.g., r_dis_obj_tar weight 3, terminal +10, dual-contact bonus +0.3
    Hand-tuned during development; they determine the learned pushing strategy and the success rate reported in Table VI.
  • Controller reward weights = Table III
    Hand-tuned to balance tracking, smoothness, and torque; affects the sim-to-real behavior used in all results.
  • Command bounds for Planner during training = v_cmd x [0,0.5] m/s, y [-0.3,0.3] m/s, yaw [-1,1] rad/s; forefoot offsets x [-0.25,0.25], y [-0.1,0.2], z [-0.2,0.2] m
    Hand-chosen bounds shape the feasible strategy space; changing them would change achievable pushing behavior.
  • Success and reset thresholds = success at 0.05 m / 5 deg / 25 s; tilt reset at 40 deg
    The same thresholds are used for the terminal reward during training and for evaluation, so they are part of the definition of the reported 95.6% success rate.
  • Domain randomization ranges = Tables IV and V
    Chosen to span expected real-world variation; the zero-shot transfer claim depends on these ranges being representative.
axioms (5)
  • domain assumption Simulation fidelity: Isaac Gym contact dynamics and PD control, with randomized physical parameters, are close enough to Daisy's real SEA-driven dynamics for zero-shot sim-to-real transfer.
    Invoked in Section III.C.1 and validated only by the qualitative real-world trials; if false, the central claim fails.
  • domain assumption Ground-truth object and robot poses are available at 50 Hz from an external motion capture system during deployment.
    Section IV.B.1 says this is the only source of object pose; the conclusion acknowledges it limits deployment to controlled environments.
  • domain assumption The low-level Controller can faithfully execute Planner commands, so freezing it and training Planner on top is a valid decomposition.
    Section III.C.2 freezes Controller parameters during Planner training; any tracking failure is assumed to be handled by the higher level.
  • domain assumption PPO with the specified rewards converges to a policy whose success reflects task competence rather than reward hacking.
    PPO is used as a black-box optimizer (Section III.C); standard practice in this literature, but no convergence guarantee is provided.
  • domain assumption Privileged object properties (mass, COM, inertia, friction) can be given only to the critic, while the actor must generalize over the randomized distribution.
    Section III.A.1; this requires the training distribution to cover real objects, and the real tests to lie within it.

pith-pipeline@v1.3.0-alltime-deepseek · 10707 in / 13858 out tokens · 290043 ms · 2026-08-04T14:38:06.859310+00:00 · methodology

0 comments
read the original abstract

In nature, animals often need to move/manipulate objects comparable in weight/size to their own bodies. Compared to grasping and carrying, pushing provides a more straightforward and efficient non-prehensile manipulation strategy, avoiding complex grasp design while leveraging direct contact to regulate an object's pose during interaction. Achieving effective pushing, however, requires both sufficient manipulation capability and stable whole-body coordination, which is particularly challenging when dealing with heavy or irregular objects. To address these challenges, we propose HeLoM, a learning-based hierarchical whole-body manipulation framework for hexapod robots that exploits coordinated multi-limb control and is applicable to multi-legged robotic systems. Inspired by the cooperative strategies of multi-legged insects, our framework leverages multiple contact points and high degrees of freedom to enable efficient and dynamic whole-body coordination during object interaction. HeLoM's high-level planner plans pushing behaviors, while its low-level controller maintains locomotion stability and generates dynamically consistent joint actions. This design enables the robot to maintain balance while executing continuous and controllable pushing behaviors through coordinated foreleg interaction and supportive hind-leg propulsion. We validate the effectiveness of HeLoM through both simulation and real-world experiments. Results show that our framework can stably push objects of varying sizes and unknown physical properties to designated goal poses in the real world.

Figures

Figures reproduced from arXiv: 2509.23651 by Arhaan Jain, Ge Sun, Guillaume Sartoretti, Hongyi Li, Junkai Lu, Linnan Chang, Peizhuo Li, Xinrong Yang, Yifeng Peng, Yifeng Zhang, Yuhong Cao.

Figure 1
Figure 1. Figure 1: Examples of the modular hexapod robot pushes different objects [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed HeLoM framework. Dashed arrows indicate components used only in simulation, while solid arrows represent components used in real-world deployment. During training, we first pre-train the Controller with randomly sampled commands to establish loco-manipulation ability. Subsequently, the network parameters of Controller are frozen, and Planner is trained on top of it to learn effecti… view at source ↗
Figure 3
Figure 3. Figure 3: Three representative pushing scenarios in simulation. A and D [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Tracking performance of HeLoM and HeLoM w/o smoothness loss with respect to desired velocity and foot position, specifically illustrating tracking of velocity in the x-axis and position in the z-axis. of HeLoM in both simulation and real-world experiments. The studies focus on two key aspects: the performance of the loco-manipulation Controller and the effectiveness of Planner in achieving pushing behavior… view at source ↗
Figure 5
Figure 5. Figure 5: Complete process of robot–object interaction in an experimental [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Four representative scenarios in real-world tests. The figure [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Human intervention introduced during the robot pushing task. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] 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

28 extracted references · 8 linked inside Pith

  1. [1]

    Legged robots for object manipulation: A review,

    Y . Gong, G. Sun, A. Nair, A. Bidwai, R. CS, J. Grezmak, G. Sartoretti, and K. A. Daltorio, “Legged robots for object manipulation: A review,” Frontiers in Mechanical Engineering, vol. 9, p. 1142421, 2023

    2023

  2. [2]

    Learning goal-oriented non-prehensile pushing in cluttered scenes,

    N. Dengler, D. Großklaus, and M. Bennewitz, “Learning goal-oriented non-prehensile pushing in cluttered scenes,” in2022 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 1116–1122

    2022

  3. [3]

    Learning whole-body manipulation for quadrupedal robot,

    S. Jeon, M. Jung, S. Choi, B. Kim, and J. Hwangbo, “Learning whole-body manipulation for quadrupedal robot,”IEEE Robotics and Automation Letters, vol. 9, no. 1, pp. 699–706, 2023

    2023

  4. [4]

    Sata: Safe and adaptive torque- based locomotion policies inspired by animal learning,

    P. Li, H. Li, G. Sun, J. Cheng, X. Yang, G. Bellegarda, M. Shafiee, Y . Cao, A. Ijspeert, and G. Sartoretti, “Sata: Safe and adaptive torque- based locomotion policies inspired by animal learning,”arXiv preprint arXiv:2502.12674, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Whole-body end- effector pose tracking,

    T. Portela, A. Cramariuc, M. Mittal, and M. Hutter, “Whole-body end- effector pose tracking,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 11 205–11 211

    2025

  6. [6]

    Dynamic object goal pushing with mobile manipulators through model-free constrained reinforcement learning,

    I. Dadiotis, M. Mittal, N. Tsagarakis, and M. Hutter, “Dynamic object goal pushing with mobile manipulators through model-free constrained reinforcement learning,”arXiv preprint arXiv:2502.01546, 2025

    arXiv 2025

  7. [7]

    Rambo: Rl- augmented model-based whole-body control for loco-manipulation,

    J. Cheng, D. Kang, G. Fadini, G. Shi, and S. Coros, “Rambo: Rl- augmented model-based whole-body control for loco-manipulation,” IEEE Robotics and Automation Letters, 2025

    2025

  8. [8]

    Nature’s all-in-one: Multi- tasking robots inspired by dung beetles,

    B. Leung, S. Gorb, and P. Manoonpong, “Nature’s all-in-one: Multi- tasking robots inspired by dung beetles,”Advanced Science, vol. 11, no. 47, p. 2408080, 2024

    2024

  9. [9]

    Contact optimization for non-prehensile loco-manipulation via hierarchical model predictive control,

    A. Rigo, Y . Chen, S. K. Gupta, and Q. Nguyen, “Contact optimization for non-prehensile loco-manipulation via hierarchical model predictive control,”arXiv preprint arXiv:2210.03442, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    Hierarchical adaptive loco- manipulation control for quadruped robots,

    M. Sombolestan and Q. Nguyen, “Hierarchical adaptive loco- manipulation control for quadruped robots,”arXiv preprint arXiv:2209.13145, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Loco-manipulation control for arm- mounted quadruped robots: Dynamic and kinematic strategies,

    G. Xin, F. Zeng, and K. Qin, “Loco-manipulation control for arm- mounted quadruped robots: Dynamic and kinematic strategies,”Ma- chines, vol. 10, no. 8, p. 719, 2022

    2022

  12. [12]

    Roloma: Robust loco-manipulation for quadruped robots with arms,

    H. Ferrolho, V . Ivan, W. Merkt, I. Havoutis, and S. Vijayakumar, “Roloma: Robust loco-manipulation for quadruped robots with arms,” Autonomous Robots, vol. 47, no. 8, pp. 1463–1481, 2023

    2023

  13. [13]

    Multi-contact heavy object pushing with a centaur-type humanoid robot: Planning and control for a real demonstrator,

    M. P. Polverini, A. Laurenzi, E. M. Hoffman, F. Ruscelli, and N. G. Tsagarakis, “Multi-contact heavy object pushing with a centaur-type humanoid robot: Planning and control for a real demonstrator,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 859–866, 2020

    2020

  14. [14]

    Dynamic locomotion in the mit cheetah 3 through convex model-predictive control,

    J. Di Carlo, P. M. Wensing, B. Katz, G. Bledt, and S. Kim, “Dynamic locomotion in the mit cheetah 3 through convex model-predictive control,” in2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2018, pp. 1–9

    2018

  15. [15]

    Dribblebot: Dynamic legged manipulation in the wild,

    Y . Ji, G. B. Margolis, and P. Agrawal, “Dribblebot: Dynamic legged manipulation in the wild,”arXiv preprint arXiv:2304.01159, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Deep whole-body control: learning a unified policy for manipulation and locomotion,

    Z. Fu, X. Cheng, and D. Pathak, “Deep whole-body control: learning a unified policy for manipulation and locomotion,” inConference on Robot Learning. PMLR, 2023, pp. 138–149

    2023

  17. [17]

    Learning force con- trol for legged manipulation,

    T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal, “Learning force con- trol for legged manipulation,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 15 366–15 372

    2024

  18. [18]

    Umi on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,

    H. Ha, Y . Gao, Z. Fu, J. Tan, and S. Song, “Umi on legs: Making manipulation policies mobile with manipulation-centric whole-body controllers,”arXiv preprint arXiv:2407.10353, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Joint-space cpg for safe foothold plan- ning and body pose control during locomotion and climbing,

    G. Sun and G. Sartoretti, “Joint-space cpg for safe foothold plan- ning and body pose control during locomotion and climbing,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 9889–9896, 2022

    2022

  20. [20]

    Autonomous obstacle legipulation with a hexapod robot,

    B. Lu, B. Tam, and N. Kottege, “Autonomous obstacle legipulation with a hexapod robot,”arXiv preprint arXiv:2011.06227, 2020

    Pith/arXiv arXiv 2011

  21. [21]

    Design and multimodal locomotion plan of a hexapod robot with improved knee joints,

    K. Xu, R. Qin, C. Chen, G. Dong, J. Chen, and X. Ding, “Design and multimodal locomotion plan of a hexapod robot with improved knee joints,”Journal of Field Robotics, vol. 41, no. 5, pp. 1279–1296, 2024

    2024

  22. [22]

    De- sign and control of a novel leg-arm multiplexing mobile operational hexapod robot,

    T. Chen, Y . Li, X. Rong, G. Zhang, H. Chai, J. Bi, and Q. Wang, “De- sign and control of a novel leg-arm multiplexing mobile operational hexapod robot,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 382–389, 2021

    2021

  23. [23]

    Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,

    P. Arm, M. Mittal, H. Kolvenbach, and M. Hutter, “Pedipulate: Enabling manipulation skills using a quadruped robot’s leg,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5717–5723

    2024

  24. [24]

    Learning multi-agent loco-manipulation for long-horizon quadrupedal pushing,

    Y . Feng, C. Hong, Y . Niu, S. Liu, Y . Yang, and D. Zhao, “Learning multi-agent loco-manipulation for long-horizon quadrupedal pushing,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 14 441–14 448

    2025

  25. [25]

    Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning,

    I. Nahrendra, B. Yu, and H. Myung, “Dreamwaq: Learning robust quadrupedal locomotion with implicit terrain imagination via deep reinforcement learning,”arXiv preprint arXiv:2301.10602, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Proximal policy optimization algorithms,

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

    Pith/arXiv arXiv 2017

  27. [27]

    L2c2: Locally lipschitz continuous constraint towards stable and smooth reinforcement learning,

    T. Kobayashi, “L2c2: Locally lipschitz continuous constraint towards stable and smooth reinforcement learning,” in2022 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 4032–4039

    2022

  28. [28]

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

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

    2022