Pith. sign in

REVIEW 3 major objections 5 minor 53 references

High-speed control and navigation for quadrupedal robots on complex and discrete terrain

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hierarchical learned tracker and sampling planner let a quadruped run on walls, clear a 1.3 m gap, and cross stepping stones at 4 m/s.

desk verdict Strong hardware results and a novel co-evolved curriculum, but the 'autonomous navigation' claim overstates what the Vicon-and-precomputed-map experiments actually demonstrate. read the letter →

arxiv 2506.02835 v1 pith:DHCLHMS4 submitted 2025-06-03 cs.RO

Hyeongjun Kim , Hyunsik Oh , Jeongsoo Park , Yunho Kim , Donghoon Youm , Moonkyu Jung , Minho Lee , Jemin Hwangbo This is my paper
classification cs.RO
keywords quadrupedallocomotionfootholdplanningreinforcementlearninggenerativecurriculumparkourdiscreteterrainhierarchicalnavigationRaibo
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 aims to show that a quadruped can navigate complex, discrete terrain at high speed by splitting the problem into two pieces: a foothold planner that proposes physically feasible landing spots, and a reinforcement-learned tracker that hits those spots accurately. The authors argue that the planner can run in real time because it samples only foothold positions, filters candidates with cheap heuristics and a learned boundary estimator, and then validates a handful of candidates with full-physics rollouts. The tracker can reach extreme agility because it is trained on terrain distributions produced by a generative model that competes with the tracker, steadily raising difficulty inside the robot's physical limits. If correct, the pipeline turns a 27 kg robot into a platform that can jump a gap 2.7 times its leg length, run on vertical walls, and follow planned routes through stairs, ramps, and stepping stones. The authors validate the claim on the Raibo quadruped, both in simulation and in hardware.

What carries the argument

The load-bearing structure is a two-module pipeline plus a training curriculum. The planner's key object is the sequential filter chain: a performance filter that checks whether a sampled foothold and its surroundings fall inside the tracker's trained range, a spike filter that uses PCA to reject rough terrain, and a collision filter that compares a learned lowest-collision-boundary estimate with the terrain height, followed by eight parallel full-physics rollouts to pick the best plan. The tracker's key object is a reinforcement-learned policy whose observation includes the next two foothold targets per foot, trained with a map generator (the decoder of a conditional variational autoencoder) that is retrained on the feasible parameter set whenever the tracker overcomes 9.3 of 10 stepping stones, adjusting a variance parameter to keep difficulty near the edge. The boundary estimator, a single MLP that predicts the lowest point swept by the robot's collision bodies between targets, is what lets the planner reject collision-prone footholds without expensive simulation.

What would settle it

Run the same planner and tracker on the same stepping-stone course with the robot's pose estimated onboard instead of by motion capture, and with a height map built on the fly from cameras; if the robot cannot maintain 4 m/s or misses footholds, then the autonomous-navigation claim, as opposed to the control and planning modules, is falsified.

Watch

Extended reading notes

Core claim

The central claim is that high-speed legged navigation on discrete terrain becomes tractable when planning and control are separated and each is made fast in the right way. The planner reduces the search to foothold positions only — rear feet step where front feet have stepped — and rejects bad samples with three sequential filters: a performance filter that checks the tracker's trained foothold range, a spike filter that rejects rough regions, and a collision filter built on a supervised boundary-estimator network. The survivors are rolled out in physics simulation with the trained tracker and the full robot model, and the lowest-cost plan is sent to the robot. The tracker is a PPO-trained policy with a GRU state estimator, trained against a map generator, a conditional-variational-autoencoder decoder that proposes distributions of stepping-stone parameters. Whenever the tracker's success crosses a threshold, the generator is retrained on the parameters the tracker overcame, keeping the curriculum at the edge of feasibility. The authors report that this adversary-like training expanded the terrain distribution to 1.6 m gaps, 60-degree step inclinations, and 90-degree wall tilts, and that the resulting system cleared a 1.3 m gap in hardware, ran on walls, crossed stepping stones at 4 m/s, and planned routes with a minimum 6.35x real-time safety factor.

Load-bearing premise

The load-bearing premise is that the robot's pose and the terrain map are provided externally through a motion-capture system and pre-acquired height maps, because onboard state estimation drifted at 4 m/s and many footholds were occluded, so the demonstrated autonomous navigation depends on inputs the system does not yet perceive itself.

Editorial extensions

If this is right

  • A single onboard computer can replan footholds faster than the control loop demands; the measured minimum safety factor across tested maps was 6.35, and the average was 18.41.
  • The tracker retains more than 90 percent simulated landing success for map height errors between -5 cm and 12 cm, and it passed five-of-five hardware trials under foot slip, movable footholds, and 15 cm map error.
  • User-defined cost functions covering survival, distance, direction, and elevation let the same planner produce different routes, such as detour versus direct paths, on identical maps.
  • Training against a generative map curriculum expands the controller's reachable terrain to gap distances and wall tilts far beyond a hand-tuned curriculum, as shown by wall-running in simulation and hardware.
  • The rear-foot-on-front-foot stepping rule and foothold-only planning bound the search space, enabling real-time operation at the cost of planning only 2.5D terrains, so vertical-wall routes are not yet planable.

Reading between the lines

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

  • If the planner accepted a 3D voxel map instead of a 2.5D height map, the same filter-and-rollout scheme could in principle plan wall-running routes; the authors explicitly leave this as future work.
  • The competitive map generator is a transferable idea: any task whose difficulty is parameterized by a low-dimensional distribution of environments could use the same retrain-when-succeed curriculum, including manipulation or whole-body tasks.
  • Because the planner relies on rear feet stepping where front feet stepped, robustness to foot slip and map error is bought in part by this redundancy; a stricter replanning policy might be needed for non-repeating footholds on highly irregular terrain.
  • A direct stress test suggested by the paper is to run the full pipeline with only onboard sensors and real-time mapping; if it fails, the gap is perception and state estimation, not the planner or tracker module.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript presents a hierarchical locomotion pipeline for quadrupedal robots on discrete and complex terrain. The tracker is an RL policy trained with a competitive CVAE-based map generator, while the planner samples foothold candidates, filters them through performance, spike, and collision filters, and rolls out eight candidate plans in simulation to select the lowest-cost plan. Hardware experiments on the Raibo robot demonstrate wall running, a 1.3 m gap jump, 4 m/s stepping-stone traversal, and real-time replanning on obstacle courses; simulated ablations and robustness trials support the tracker and planner components.

Significance. The paper has substantial strengths: real-robot repetitions with 5/5 success in slip, unstable foothold, and map-height-error scenarios; 500-trial simulated ablations over five maps; a minimum planning safety factor of 6.35; and detailed supplementary material on network architectures, rewards, and curricula. The competitive generative-model curriculum is a plausible contribution for training high-performance trackers in high-dimensional terrain-parameter spaces. The main weakness is that the headline claim of 'autonomous navigation' is not supported by the deployed system, which uses pre-acquired height maps and external Vicon pose. If the claims are reframed to describe offline-planned foothold tracking with external localization, the central algorithmic and experimental contributions remain significant and defensible.

major comments (3)
  1. [Abstract; Experimental setup] The abstract and contributions use 'autonomously navigating' and 'autonomous navigation,' but the Experimental setup explicitly states that a perception system was not used, height maps were pre-acquired, and the robot's pose came from a Vicon motion-capture system. The system that was actually demonstrated is a high-speed foothold tracker with an offline foothold planner and external localization, not an autonomous navigation system. Please revise the Abstract, Results, and Discussion to use precise wording such as 'semi-autonomous' or 'offline-planned navigation with external state estimation,' and state the external inputs in the claims. This is load-bearing because the headline claim as written overstates what the experiments demonstrate.
  2. [Experimental setup; Input representation for Actor and State Estimator] The sentence 'The robot's pose was retrieved only at the moment when the planner module was activated' is ambiguous and potentially inconsistent with the rest of the pipeline. The tracker must convert world-frame foothold targets into body-frame observations at each target-index update, which requires knowledge of the robot's global pose at those times. Please specify exactly when external pose is used and how the tracker obtains the foothold targets in body frame during deployment; otherwise the architecture description is incomplete.
  3. [Related work; Evaluation of tracker] The Abstract claims that 'The resulting tracker can overcome terrains that are more difficult than what the previous methods could manage,' but the only quantitative prior-robot comparisons are gap distances and leg-length ratios to Anymal D and Unitree A1. These examples do not establish a general difficulty ordering, and the leg-length values used for the ratios should be cited explicitly. Please either restrict the claim to the demonstrated scenarios or add a systematic comparison with prior systems under matched conditions.
minor comments (5)
  1. [Table 1] The text says the safety factor is 'calculated by T_update dividing the total time consumed,' but the table values correspond to T_update divided by total planning time; please correct the wording to avoid ambiguity.
  2. [Statistical analysis, Eq. (1)] Equation (1) defines the population standard deviation, not the sample standard deviation; please label it accordingly and clarify which quantity is used in Fig. 3D.
  3. [Evaluation of planner] The boundary estimator is reported with an RMSE of 2.27 cm and a 95% error bound, but no dataset split or test-set description is given; please specify the training/test procedure and the number of samples.
  4. [Training procedure; Algorithm S1] The performance thresholds 9.3 and 9.15 and the alpha decrement of 0.02 are hand-tuned; please comment on their sensitivity and how they were selected, since the competitive-training claim depends on these values.
  5. [Planner module] The performance filter is described as using a 'carefully tuned box constraint,' but no numerical values or tuning procedure are provided; please add the box dimensions and any supporting sensitivity analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the headline results are empirical hardware and simulation outcomes, and the map-generator/planner closed loop is a legitimate adaptive-curriculum and closed-loop design, not a fit renamed as a prediction.

full rationale

The paper's claimed achievements (1.3 m gap, 4 m/s stepping stones, wall runs, planner safety factor at least 6.35) are measured outcomes from real-robot trials and simulation rollouts, not quantities derived from fitted constants or from assumptions containing the conclusions. The map generator is trained competitively: it is retrained on foothold sequences the tracker has already overcome, and alpha is adjusted to keep difficulty near the tracker's current level (Algorithm S1). This is an adaptive curriculum; the statement that the tracker is trained at the 'desired difficulty' is a definition of the curriculum, not a derivation of the headline capability. Likewise, the planner's performance filter restricts samples to the tracker's training range and candidate plans are then re-simulated with the same tracker (Fig. 2, Algorithm S2); evaluating a controller-planner pair in closed loop is not circular because success is measured on separately defined maps and real scenarios. Self-citations (Raibo hardware, the Raisim simulator, and the state-estimator architecture) provide platform, tools, and design choices but do not carry the load of the claimed result, which is demonstrated on hardware and compared against independent prior work. The acknowledged reliance on Vicon pose and pre-acquired height maps is a limitation on the 'autonomous navigation' claim and a correctness/scope concern, but it is not a circularity: it does not make any predicted quantity equal to its input by construction.

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

The system is largely empirical: many hand-tuned constants and a simulated training environment underpin both the learned tracker and the planner. The central claims rest on simulation fidelity and externally supplied state/terrain information. No new physical entities are introduced.

free parameters (6)
  • Initial curriculum ranges (Stage 0-4 for r, theta, phi, dyaw, xtilt, ytilt) = table S3 values, e.g., r_low=0.4 m, r_high=0.8 to 1.1 m, theta=5 to 20 deg, phi=5 to 40 deg, xtilt=10 to 30 deg
    Hand-specified schedule used to grow terrain difficulty in the first stage of training (Methods, Terrain generation).
  • Adversarial training thresholds and alpha schedule = performance threshold 9.3/9.15 stones; alpha starts at 0.7 and decrements by 0.02
    Algorithm 1; these values determine when the map generator is retrained and how hard the generated maps are. They are manually tuned and no sensitivity analysis is given.
  • Reward coefficients in table S4 = e.g., k_ts1=9.4, k_ts2=0.97, k_td=0.30, k_bound=1, k_impact=-2
    All reward weights are hand-selected during training; the paper does not report a sweep or ablations over these coefficients.
  • Planner design constants = 8 candidate plans, saferadius=6 cm, num_target=2, performance-filter box constraint
    Methods/Planner module and Supplementary Methods state these were determined through simulations and real experiments to balance stability and time complexity.
  • Sampling-range polynomial models = Coefficients learned by supervised regression of r, theta, dyaw ranges conditioned on previous psi
    Instead of networks, the planner models sampling ranges as polynomials fit to data (Supplementary Methods); the coefficients are fitted but their values are not reported.
  • Planning cost coefficients = k_survive=-10000, k_distance=1, k_direction=0.0333, k_elevation=1
    Table S5; these weights shape plan selection and were chosen by the authors.
assumptions (5)
  • domain assumption Raisim contact dynamics simulation accurately models Raibo's behavior for both training and online plan evaluation.
    The tracker is trained entirely in Raisim (Methods, Training procedure) and the planner selects foothold plans by rolling out the trained tracker in the same simulator (Methods, Planner module). If simulation diverges from the real robot, plan feasibility and the learned policy would not transfer.
  • domain assumption The 6D terrain parameter psi and the CVAE map generator can represent the full range of deployment terrain (including 90-degree wall segments).
    Terrain generation and curriculum evolution (Methods, Terrain generation) are defined on psi; the planner's foothold search is also bounded by distributions derived from this parameterization. The final demonstration terrains (walls, 1.3 m gaps, 4 m/s stepping stones) are assumed to lie inside the generated distribution.
  • domain assumption External motion capture (Vicon) and pre-acquired height maps provide the state and terrain information needed for real deployment.
    The authors state onboard state estimation was too inaccurate for high-speed movement and that footholds were occluded, so they used Vicon pose and pre-acquired height maps (Experimental setup). This is an assumption about available infrastructure, not a demonstrated onboard capability.
  • domain assumption The learned boundary estimator generalizes from its simulated training data to unseen terrain and real conditions.
    The collision filter uses a single MLP trained via supervised learning in simulation (Methods, Planner module). Its 2.27 cm RMSE is reported only on training-like data, not on real sensor or state-estimation noise.
  • ad hoc to paper Hand-tuned curriculum thresholds (9.3/9.15 performance) and alpha adjustment produce a terrain distribution that remains feasible for the robot.
    Algorithm 1 sets alpha=0.7 and decrements by 0.02 based on the tracker's average stones overcome; these values and the retraining threshold are chosen to stabilize adversarial training and are not derived from a formal guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-speed control and navigation for quadrupedal robots on complex and discrete terrain." pith.science (2026). https://pith.science/paper/DHCLHMS4

@misc{pith2026250602835,
  author       = {Pith},
  title        = {Pith review of: High-speed control and navigation for quadrupedal robots on complex and discrete terrain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHCLHMS4}},
  note         = {Machine review of arXiv:2506.02835}
}
read the original abstract

High-speed legged navigation in discrete and geometrically complex environments is a challenging task because of the high-degree-of-freedom dynamics and long-horizon, nonconvex nature of the optimization problem. In this work, we propose a hierarchical navigation pipeline for legged robots that can traverse such environments at high speed. The proposed pipeline consists of a planner and tracker module. The planner module finds physically feasible foothold plans by sampling-based optimization with fast sequential filtering using heuristics and a neural network. Subsequently, rollouts are performed in a physics simulation to identify the best foothold plan regarding the engineered cost function and to confirm its physical consistency. This hierarchical planning module is computationally efficient and physically accurate at the same time. The tracker aims to accurately step on the target footholds from the planning module. During the training stage, the foothold target distribution is given by a generative model that is trained competitively with the tracker. This process ensures that the tracker is trained in an environment with the desired difficulty. The resulting tracker can overcome terrains that are more difficult than what the previous methods could manage. We demonstrated our approach using Raibo, our in-house dynamic quadruped robot. The results were dynamic and agile motions: Raibo is capable of running on vertical walls, jumping a 1.3-meter gap, running over stepping stones at 4 meters per second, and autonomously navigating on terrains full of 30{\deg} ramps, stairs, and boxes of various sizes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 41 canonical work pages

  1. [1]

    B. Katz, J. Di Carlo, S. Kim, Mini cheetah: A platform for pushing the limits of dynamic quadruped control, in2019 international conference on robotics and automation (ICRA)(IEEE) (2019), pp. 6295–6301

    work page 2019

  2. [2]

    Hutter,et al., Anymal-a highly mobile and dynamic quadrupedal robot, in2016 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE) (2016), pp

    M. Hutter,et al., Anymal-a highly mobile and dynamic quadrupedal robot, in2016 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE) (2016), pp. 38–44

    work page 2016

  3. [3]

    Y.-H. Shin,et al., Design of KAIST HOUND, a quadruped robot platform for fast and efficient locomotion with mixed-integer nonlinear optimization of a gear train, in2022 International Conference on Robotics and Automation (ICRA)(IEEE) (2022), pp. 6614–6620

    work page 2022

  4. [4]

    Semini,et al., Design of HyQ–a hydraulically and electrically actuated quadruped robot

    C. Semini,et al., Design of HyQ–a hydraulically and electrically actuated quadruped robot. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering225(6), 831–849 (2011)

    work page 2011

  5. [5]

    Miki,et al., Learning robust perceptive locomotion for quadrupedal robots in the wild.Sci

    T. Miki,et al., Learning robust perceptive locomotion for quadrupedal robots in the wild.Sci. Robot.7(62), eabk2822 (2022)

    work page 2022

  6. [6]

    Shi,et al., Rethinking Robustness Assessment: Adversarial Attacks on Learning-based Quadrupedal Locomotion Controllers.arXiv:2405.12424(2024)

    F. Shi,et al., Rethinking Robustness Assessment: Adversarial Attacks on Learning-based Quadrupedal Locomotion Controllers.arXiv:2405.12424(2024)

    arXiv 2024

  7. [7]

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

    arXiv 2024

  8. [8]

    Jenelten, R

    F. Jenelten, R. Grandia, F. Farshidian, M. Hutter, TAMOLS: Terrain-aware motion optimization for legged systems.T-RO38(6), 3395–3413 (2022)

    work page 2022

Show all 53 references
  1. [9]

    Abdalla, M

    A. Abdalla, M. Focchi, R. Orsolino, C. Semini, An Efficient Paradigm for Feasibility Guarantees in Legged Locomotion.T-RO39(5), 3499–3515 (2023)

    2023

  2. [10]

    Di Carlo, P

    J. Di Carlo, P. M. Wensing, B. Katz, G. Bledt, 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

  3. [11]

    D. Kim, J. Di Carlo, B. Katz, G. Bledt, S. Kim, Highly dynamic quadruped locomotion via whole-body impulse control and model predictive control.arXiv:1909.06586(2019)

    2019 arXiv

  4. [12]

    H. Chen, Z. Hong, S. Yang, P. M. Wensing, W. Zhang, Quadruped capturability and push recovery via a switched-systems characterization of dynamic balance.IEEE T-RO39(3), 2111–2130 (2023)

    2023

  5. [13]

    M. V. Minniti, R. Grandia, F. Farshidian, M. Hutter, Adaptive CLF-MPC with application to quadrupedal robots.IEEE Robot. Autom. Lett.7(1), 565–572 (2021)

    2021

  6. [14]

    H. Li, P. M. Wensing, Cafe-mpc: A cascaded-fidelity model predictive control framework with tuning-free whole-body control.arXiv:2403.03995(2024). 22

    2024 arXiv

  7. [15]

    Garc ´ıa, R

    G. Garc ´ıa, R. Griffin, J. Pratt, Time-varying model predictive control for highly dynamic motions of quadrupedal robots, in2021 IEEE International Conference on Robotics and Automation (ICRA)(IEEE) (2021), pp. 7344–7349

    2021

  8. [16]

    Hong, J.-H

    S. Hong, J.-H. Kim, H.-W. Park, Real-time constrained nonlinear model predictive control on so (3) for dynamic legged locomotion, in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2020), pp. 3982–3989

    2020

  9. [17]

    Nguyen, L

    C. Nguyen, L. Bao, Q. Nguyen, Continuous jumping for legged robots on stepping stones via trajectory optimization and model predictive control, in2022 IEEE 61st Conference on Decision and Control (CDC)(IEEE) (2022), pp. 93–99

    2022

  10. [18]

    Grandia, F

    R. Grandia, F. Jenelten, S. Yang, F. Farshidian, M. Hutter, Perceptive locomotion through nonlinear model-predictive control.IEEE T-RO39(5), 3402–3421 (2023)

    2023

  11. [19]

    Agrawal, S

    A. Agrawal, S. Chen, A. Rai, K. Sreenath, Vision-aided dynamic quadrupedal locomotion on discrete terrain using motion libraries, in2022 International Conference on Robotics and Automation (ICRA)(IEEE) (2022), pp. 4708–4714

    2022

  12. [20]

    Choi,et al., Learning quadrupedal locomotion on deformable terrain.Sci

    S. Choi,et al., Learning quadrupedal locomotion on deformable terrain.Sci. Robot.8(74), eade2256 (2023)

    2023

  13. [21]

    G. Ji, J. Mun, H. Kim, J. Hwangbo, Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion.IEEE Robot. Autom. Lett.7(2), 4630–4637 (2022)

    2022

  14. [22]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, J. Malik, Rma: Rapid motor adaptation for legged robots. arXiv:2107.04034(2021)

    2021 arXiv

  15. [23]

    Feng,et al., Genloco: Generalized locomotion controllers for quadrupedal robots, inCon- ference on Robot Learning(PMLR) (2023), pp

    G. Feng,et al., Genloco: Generalized locomotion controllers for quadrupedal robots, inCon- ference on Robot Learning(PMLR) (2023), pp. 1893–1903

    2023

  16. [24]

    X. B. Peng, M. Andrychowicz, W. Zaremba, P. Abbeel, Sim-to-real transfer of robotic con- trol with dynamics randomization, in2018 IEEE international conference on robotics and automation (ICRA)(IEEE) (2018), pp. 3803–3810

    2018

  17. [25]

    Kim,et al., Not only rewards but also constraints: Applications on legged robot locomotion

    Y. Kim,et al., Not only rewards but also constraints: Applications on legged robot locomotion. IEEE T-RO(2024)

    2024

  18. [26]

    Jenelten, J

    F. Jenelten, J. He, F. Farshidian, M. Hutter, DTC: Deep Tracking Control.Sci. Robot.9(86), eadh5401 (2024)

    2024

  19. [27]

    Hoeller, N

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

    2024

  20. [28]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, D. Pathak, Extreme parkour with legged robots. arXiv:2309.14341(2023)

    2023 arXiv

  21. [29]

    Duan,et al., Learning dynamic bipedal walking across stepping stones, in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp

    H. Duan,et al., Learning dynamic bipedal walking across stepping stones, in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 6746– 6752. 23

    2022

  22. [30]

    Rudin, D

    N. Rudin, D. Hoeller, M. Bjelonic, M. Hutter, Advanced skills by learning locomotion and local navigation end-to-end, in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 2497–2503

    2022

  23. [31]

    S. Jeon, M. Jung, S. Choi, B. Kim, J. Hwangbo, Learning whole-body manipulation for quadrupedal robot.IEEE Robot. Autom. Lett.9(1), 699–706 (2023)

    2023

  24. [32]

    R. Wang, J. Lehman, J. Clune, K. O. Stanley, Paired open-ended trailblazer (poet): End- lessly generating increasingly complex and diverse learning environments and their solutions. arXiv:1901.01753(2019)

    2019 arXiv

  25. [33]

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

    2020

  26. [34]

    Z. Xie, H. Y. Ling, N. H. Kim, M. van de Panne, Allsteps: curriculum-driven learning of stepping stone skills, inComputer Graphics Forum(Wiley Online Library), vol. 39 (2020), pp. 213–224

    2020

  27. [35]

    Lee,et al., Learning robust autonomous navigation and locomotion for wheeled-legged robots.Sci

    J. Lee,et al., Learning robust autonomous navigation and locomotion for wheeled-legged robots.Sci. Robot.9(89), eadi9641 (2024)

    2024

  28. [36]

    Zhang,et al., Resilient legged local navigation: Learning to traverse with compromised perception end-to-end, in41st IEEE Conference on Robotics and Automation (ICRA 2024) (2024)

    C. Zhang,et al., Resilient legged local navigation: Learning to traverse with compromised perception end-to-end, in41st IEEE Conference on Robotics and Automation (ICRA 2024) (2024)

    2024

  29. [37]

    Y. Kim, C. Kim, J. Hwangbo, Learning forward dynamics model and informed trajectory sampler for safe quadruped navigation.arXiv:2204.08647(2022)

    2022 arXiv

  30. [38]

    Chestnutt,Navigation planning for legged robots(Carnegie Mellon University) (2007)

    J. Chestnutt,Navigation planning for legged robots(Carnegie Mellon University) (2007)

    2007

  31. [39]

    Wellhausen, M

    L. Wellhausen, M. Hutter, Rough terrain navigation for legged robots using reachability plan- ning and template learning, in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2021), pp. 6914–6921

    2021

  32. [40]

    Xu,et al., Contact sequence planning for hexapod robots in sparse foothold environment based on Monte-Carlo tree.IEEE Robot

    P. Xu,et al., Contact sequence planning for hexapod robots in sparse foothold environment based on Monte-Carlo tree.IEEE Robot. Autom. Lett.7(2), 826–833 (2021)

    2021

  33. [41]

    Tsounis, M

    V. Tsounis, M. Alge, J. Lee, F. Farshidian, M. Hutter, Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning.IEEE Robot. Autom. Lett.5(2), 3699– 3706 (2020)

    2020

  34. [42]

    Heess,et al., Emergence of locomotion behaviours in rich environments.arXiv:1707.02286 (2017)

    N. Heess,et al., Emergence of locomotion behaviours in rich environments.arXiv:1707.02286 (2017)

    2017 arXiv

  35. [43]

    Caluwaerts,et al., Barkour: Benchmarking animal-level agility with quadruped robots

    K. Caluwaerts,et al., Barkour: Benchmarking animal-level agility with quadruped robots. arXiv:2305.14654(2023)

    2023 arXiv

  36. [44]

    X. B. Peng, G. Berseth, K. Yin, M. Van De Panne, Deeploco: Dynamic locomotion skills using hierarchical deep reinforcement learning.TOG36(4), 1–13 (2017). 24

    2017

  37. [45]

    Brakel, S

    P. Brakel, S. Bohez, L. Hasenclever, N. Heess, K. Bousmalis, Learning coordinated terrain- adaptive locomotion by imitating a centroidal dynamics planner, in2022 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 10335–10342

    2022

  38. [46]

    Mastalli, I

    C. Mastalli, I. Havoutis, A. W. Winkler, D. G. Caldwell, C. Semini, On-line and on-board planning and perception for quadrupedal locomotion, in2015 IEEE International Conference on Technologies for Practical Robot Applications (TePRA)(IEEE) (2015), pp. 1–7

    2015

  39. [47]

    F. Risbourg,et al., Real-time footstep planning and control of the solo quadruped robot in 3d environments, in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2022), pp. 12950–12956

    2022

  40. [48]

    Schulman, F

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

    2017 arXiv

  41. [49]

    Hwangbo, J

    J. Hwangbo, J. Lee, M. Hutter, Per-contact iteration method for solving contact dynamics. IEEE Robot. Autom. Lett.3(2), 895–902 (2018),www.raisim.com

    2018

  42. [50]

    D. P. Kingma, S. Mohamed, D. Jimenez Rezende, M. Welling, Semi-supervised learning with deep generative models.Advances in neural information processing systems27(2014)

    2014

  43. [51]

    Rosenblatt, The perceptron: a probabilistic model for information storage and organization in the brain.Psychological review65(6), 386 (1958)

    F. Rosenblatt, The perceptron: a probabilistic model for information storage and organization in the brain.Psychological review65(6), 386 (1958)

    1958

  44. [52]

    Cho,et al., Learning phrase representations using RNN encoder-decoder for statistical machine translation.arXiv:1406.1078(2014)

    K. Cho,et al., Learning phrase representations using RNN encoder-decoder for statistical machine translation.arXiv:1406.1078(2014)

    2014 arXiv

  45. [53]

    Deits, R

    R. Deits, R. Tedrake, Footstep planning on uneven terrain with mixed-integer convex opti- mization, in2014 IEEE-RAS international conference on humanoid robots(IEEE) (2014), pp. 279–286. 25 Supplementary Methods Network details Actor Actor is a neural network responsible for c...

    2014

Pith tools

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