Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Skill-Nav: Enhanced Navigation with Versatile Quadrupedal Locomotion via Waypoint Interface

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Skill-Nav shows that 2D waypoints, not velocity commands, let off-the-shelf planners drive a learned quadruped locomotion policy through complex terrain.

desk verdict Competent waypoint-interface navigation paper with a real-hardware demo, but the LLM prompt overstates the policy's limits and the evaluation lacks seeds. read the letter →

arxiv 2506.21853 v3 pith:ZUNBEOHY submitted 2025-06-27 cs.RO

Dewei Wang , Chenjia Bai , Chenhui Li , Jiyuan Shi , Yan Ding , Chi Zhang , Bin Zhao This is my paper
classification cs.RO
keywords ReinforcementLearningQuadrupedalRobotsVisualLocomotionwaypointinterfacehierarchicalnavigationlargelanguagemodelspathplanningsim-to-realtransfer
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 replacing velocity commands with 2D waypoints as the command input to a learned quadruped locomotion policy creates a clean interface between high-level planners and low-level control. It argues that a single waypoint-guided policy, trained in two simulated scenarios, can handle the irregular waypoints a general planner (A* or a large language model) would emit, and can climb, jump, and bypass obstacles it was never explicitly told about. The authors support this with simulation and real-robot navigation experiments on a Unitree AlienGo. If correct, the result would let off-the-shelf navigation tools drive agile legged robots using only coarse maps and a short capability description.

What carries the argument

The central object is the waypoint command $w_t$, a two-dimensional target position expressed relative to the robot's base frame, which replaces the usual velocity command in the policy observation. The argument is carried by reward functions built around waypoint tracking: $r_{\text{reach}}$ rewards the number of distant waypoints reached per time, $r_{\text{stay}}$ keeps the robot planted at a reached waypoint until the next command arrives, and a modified $r_{\text{track}}$ encourages forward motion toward the waypoint while penalizing retreat. The training pipeline is teacher-student: a teacher policy trained with privileged terrain scandots in the WP-Fixed then WP-Random scenarios is distilled into a student policy that uses only proprioception, history, and depth images, with inflated virtual obstacles injected during distillation to teach safe distances from obstacles. These pieces together make waypoints a sparse, flexible signal that high-level planners can generate from coarse maps.

What would settle it

Run the LLM planner on a map that includes a gap wider than 0.7 m or a platform higher than 0.45 m, so that the planner's stated capabilities are exceeded; if the robot falls into the gap or fails to climb and cannot recover, the claim that coarse capability descriptions suffice for waypoint generation breaks.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the waypoint is a sufficient interface: a deep-RL locomotion policy that observes only its current waypoint, a 2D position in the robot's frame, can exhibit the full repertoire of climbing, jumping, and obstacle avoidance needed to traverse complex terrain, provided it is trained sequentially in a fixed-waypoint skill-learning scenario and a random-waypoint robustness scenario. This policy, after teacher-student distillation with inflated virtual obstacles, succeeds in single-traverse and omni-traverse tasks with high obstacles where RMA and Extreme Parkour baselines fail, and it completes real-world navigation when driven by A* waypoints or GPT-4-generated waypoints. The paper's own comparison of ablations shows that neither training scenario alone is sufficient: WP-Fixed-only yields a policy that cannot follow irregular waypoints, and WP-Random-only yields a jumping, unstable gait.

Load-bearing premise

The high-level planner's description of what the robot can do must match the policy's true limits; if the planner emits a waypoint inside a gap or on a box edge, the low-level policy is not trained to recover and can get stuck.

Editorial extensions

If this is right

  • A* or Dijkstra path planners can drive the low-level policy using only an occupancy map annotated with walls, so no fine-grained elevation map or terrain classifier is needed for navigation.
  • An LLM such as GPT-4 can act as the high-level planner, generating waypoint sequences from a coarse grid map plus a short description of the robot's locomotion capabilities.
  • The low-level policy can recover from low obstacles that the depth camera fails to detect, and from external disturbances that push the robot off its path.
  • The two-scenario training schedule is essential; training in only one scenario degrades the policy into either waypoint-tracking failure or an unstable jumping gait.
  • Robots can be waypointed through terrain with obstacles not described in the plan; the policy handles them through its learned locomotion skills.

Reading between the lines

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

  • The waypoint interface should transfer to other legged platforms and possibly wheeled robots as long as the policy is retrained, since the planner only needs 2D targets.
  • In genuinely new terrains, the capability text handed to the LLM must be estimated rather than matched to the test set; otherwise the infeasible-waypoint failure the authors concede for gaps and box edges will appear regularly.
  • The inflated-virtual-obstacle distillation trick is a simple safety-enhancement module that could be tested in isolation on other depth-based locomotion policies.
  • A deployable system would benefit from a waypoint feasibility check, such as refusing waypoints whose local terrain is a gap or a steep edge, because the policy does not guarantee recovery from those placements.
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 / 4 minor

Summary. The paper proposes Skill-Nav, a hierarchical quadruped navigation framework in which a high-level planner (A* or an LLM) emits 2D waypoints and a low-level RL policy, trained in two simulated scenarios (WP-Fixed and WP-Random) and distilled to a student policy, tracks those waypoints while climbing, jumping, and bypassing obstacles. The central claim is that the waypoint interface makes the low-level policy compatible with general planning tools, so that coarse-grained terrain information suffices for complex navigation. Experiments compare the locomotion policy with RMA and Extreme Parkour in two simulated tasks and provide qualitative simulated and real-world demonstrations with GPT-4 and A*.

Significance. If the central claim holds, Skill-Nav offers a simple and modular interface that decouples high-level reasoning from low-level agility, with the practical benefit of using LLMs or classical planners without end-to-end training. The paper's strengths include a two-scenario training design, an inflated virtual obstacle trick for safer distillation, a real Unitree AlienGo deployment, and the stated availability of code from the first author. The main limitation is evidentiary: the quantitative locomotion comparison lacks statistical support, and the navigation system, which is the subject of the abstract, is demonstrated only anecdotally. I also regard the LLM capability prompt as a legitimate design choice rather than a circularity, but the capability values given to the planner must match the policy actually deployed.

major comments (3)
  1. [§3.3.2, Fig. 4, Table 1] The capability description supplied to the LLM is not consistent with the training curriculum of the deployed policy. Section 3.3.2 states that the description is 'based on the most difficult terrain the robot can traverse during the test', and the prompt in Fig. 4 tells the LLM the robot can climb platforms up to 0.45 m and jump gaps up to 0.7 m. However, Table 1 caps the WP-Random scenario, the only scenario in which the student policy is trained according to Fig. 2, at 0.35 m box height and 0.35 m gap width, and Section 4.3 states the test terrain is identical to the highest WP-Random curriculum level. The paper therefore does not establish that the final policy can execute 0.45 m climbs or 0.7 m jumps, and Section 4.4 concedes that waypoints placed in gaps or at box edges are not reliably recoverable. Because every planner variant relies on this capability model to avoid infeasible commands, the feasibility of planner-generated waypoints is a load-bearing assumption. The authors should either validate the student policy on 0.45 m/0.7 m terrain or align the prompt with empirically validated limits, and then report planner success rates on that basis.
  2. [§4.3, Table 2] The locomotion comparison is reported as single point estimates without error bars, seeds, or statistical tests, and the number of evaluation episodes per configuration is not stated beyond the initial 18 robots. Differences such as Ours (SR 0.89) versus Ours-s2 (SR 0.83) in the omni-traverse task with obstacles may be within sampling noise, yet the text interprets them as substantive evidence. In addition, the RMA baseline scores 0.00 on every task, which is surprising for a method known to walk over challenging terrain; this suggests either a mismatched evaluation protocol or an implementation issue that should be explained. The authors should report means and variances over at least three training seeds, state episode counts explicitly, and either correct or carefully justify the RMA baseline.
  3. [§4.4] The navigation system, which is the central contribution, is evaluated only qualitatively: one simulated GPT-4 demo and one real-world A* demo, with no success rates, number of trials, completion times, or comparison between planners. The abstract claims 'extensive experiments', and the waypoint interface is meant to enable 'various general planning tools'; this requires at least a small quantitative study, for example N trials per planner on several maps, reporting success, path length, and behavior when waypoints are infeasible. Without such a study, the main claim is supported only by anecdote.
minor comments (4)
  1. [§3.2.2] The description of waypoint candidate selection should define the orientation threshold precisely: it is unclear whether '≤90°' refers to the angle between the robot's heading and the waypoint direction, and whether candidates are sampled uniformly within that cone.
  2. [Declarations] The Declarations section contains an incomplete sentence, 'may be limited by the attached paper reports the research', which should be corrected or removed.
  3. [Code availability] The statement that code is 'available from the first author' is not reproducible enough; the authors should provide a public repository or a clear reason for withholding it.
  4. [Table 2] The definition of AST should state explicitly in the table caption that failed episodes are recorded as the maximum test time when SR is non-zero, and the meaning of '/' should be repeated in the caption for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: Skill-Nav is an empirical system paper with no derivation that reduces to its inputs.

full rationale

Skill-Nav is an empirical robotics paper: the low-level policy is trained with PPO, reward shaping, curriculum scenarios, and teacher-student distillation, and the high-level planner (A* or an LLM) produces waypoints that are then evaluated in simulation and on hardware. There is no analytic derivation chain in which a claimed prediction is equivalent to a fitted input by construction. The LLM prompt capability values (0.45 m platform climb, 0.7 m gap width) are hand-specified and do not match the WP-Random curriculum caps in Table 1 (0.35 m box height, 0.35 m gap width), and Section 4.4 concedes that waypoints placed in gap centers or box edges are not reliably recovered; this is a correctness and generalization risk, not circularity, because the navigation outcome is not guaranteed by construction and the planner outputs are empirically tested. The paper's citations to prior work such as [16] and [28] provide external baselines and training paradigms rather than a self-citation chain that forces the central claim. No fitted parameter is renamed as a prediction, and no target quantity is defined in terms of another target quantity. The central claim that waypoints serve as a flexible interface is supported by direct experiments rather than by definitional equivalence.

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

The central claim rests on several hand-chosen training parameters and standard RL sim-to-real assumptions rather than a mathematical derivation. The most consequential free parameter is the LLM capability description, because a mismatch between the description and the policy's real limits would make planned waypoints infeasible and break navigation.

free parameters (6)
  • Waypoint distance threshold = aligned with terrain unit size (e.g., 2m in nav tests)
    In the WP-Random scenario, waypoint candidates are restricted to positions within a distance threshold equal to the terrain unit size to prevent crossing multiple units (Sec. 3.2.2).
  • Waypoint dwell time = 2 seconds
    The target waypoint shifts after the robot remains near it for more than two seconds (Sec. 3.2.1 and 3.2.2); this hand-chosen constant shapes both training episodes and deployment behavior.
  • Reward function coefficients = not reported
    The rewards for waypoint reaching, staying, tracking, and regularization are defined in Sec. 3.2 but their relative weights are not listed; the paper defers to refs 28, 29, 38, so exact values are free parameters.
  • Terrain curriculum ranges = Table 1, e.g., box height up to 0.5m in WP-Fixed
    The ranges for hurdle height, box height, gap width, and obstacle size are hand-selected curriculum parameters that determine which skills the policy learns (Sec. 4.1).
  • LLM capability description = climb 0.45m, jump gaps 0.7m
    The LLM prompt in Sec. 3.3.2 is given the robot's capabilities as climbing platforms up to 0.45m and jumping gaps up to 0.7m, hand-set to match the most difficult test terrain.
  • Inflated virtual obstacle scale = not specified
    Obstacles are enlarged during student-policy distillation to teach safe distances, but the inflation factor is not reported (Sec. 3.2.2).
assumptions (5)
  • domain assumption Sim-to-real transfer via domain randomization and depth noise in Isaac Gym yields a deployable policy
    The deployment section assumes that a policy trained in simulation with randomized dynamics and depth-image noise transfers to the Unitree Aliengo hardware (Sec. 4.1 and 4.2).
  • domain assumption Motion capture provides accurate localization for real-world navigation
    Real-world experiments rely on a high-frequency motion capture system for localization (Sec. 4.2 and 4.4); the robot's own onboard localization is not used, so navigation performance under onboard estimation is unverified.
  • domain assumption Teacher-student distillation preserves locomotion skills under information loss
    The teacher policy receives privileged terrain scandots and encoded information, while the student policy sees only depth and proprioception; the paper assumes distillation retains the teacher's capabilities (Sec. 3.2).
  • domain assumption The high-level planner's coarse map faithfully represents the environment
    The A* planner uses an occupancy map with only wall information, and the LLM uses a height array; if the coarse map is inaccurate, planned waypoints may be infeasible, and the low-level policy is not guaranteed to recover (Sec. 3.3 and 4.4).
  • domain assumption PPO and the custom reward shaping converge to the desired locomotion behaviors
    Training relies on Proximal Policy Optimization with the tailored reward functions in Sec. 3.2; convergence and reward-weight stability are assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Skill-Nav: Enhanced Navigation with Versatile Quadrupedal Locomotion via Waypoint Interface." pith.science (2026). https://pith.science/paper/ZUNBEOHY

@misc{pith2026250621853,
  author       = {Pith},
  title        = {Pith review of: Skill-Nav: Enhanced Navigation with Versatile Quadrupedal Locomotion via Waypoint Interface},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUNBEOHY}},
  note         = {Machine review of arXiv:2506.21853}
}
read the original abstract

Quadrupedal robots have demonstrated exceptional locomotion capabilities through Reinforcement Learning (RL), including extreme parkour maneuvers. However, integrating locomotion skills with navigation in quadrupedal robots has not been fully investigated, which holds promise for enhancing long-distance movement capabilities. In this paper, we propose Skill-Nav, a method that incorporates quadrupedal locomotion skills into a hierarchical navigation framework using waypoints as an interface. Specifically, we train a waypoint-guided locomotion policy using deep RL, enabling the robot to autonomously adjust its locomotion skills to reach targeted positions while avoiding obstacles. Compared with direct velocity commands, waypoints offer a simpler yet more flexible interface for high-level planning and low-level control. Utilizing waypoints as the interface allows for the application of various general planning tools, such as large language models (LLMs) and path planning algorithms, to guide our locomotion policy in traversing terrains with diverse obstacles. Extensive experiments conducted in both simulated and real-world scenarios demonstrate that Skill-Nav can effectively traverse complex terrains and complete challenging navigation tasks.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Hume: Introducing System-2 Thinking in Visual-Language-Action Model

    cs.RO 2025-05 conditional novelty 6.0 of 10

    A dual-system vision-language-action model that improves robot control by ranking multiple sampled action chunks with a learned value function before fast execution.

Reference graph

Works this paper leans on

67 extracted references · 45 canonical work pages · cited by 1 Pith paper

  1. [1]

    Neurocomputing275, 1132–1139 (2018)

    Jiang, X., Pang, Y., Li, X., Pan, J., Xie, Y.: Deep neural networks with elas- tic rectified linear units for object recognition. Neurocomputing275, 1132–1139 (2018)

    work page 2018

  2. [2]

    IEEE transactions on neural networks and learning systems26(10), 2275–2290 (2015)

    Gao, F., Tao, D., Gao, X., Li, X.: Learning to rank for blind image quality assessment. IEEE transactions on neural networks and learning systems26(10), 2275–2290 (2015)

    work page 2015

  3. [3]

    In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), vol

    Tao, D., Li, X., Maybank, S.J., Wu, X.: Human carrying status in visual surveil- lance. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), vol. 2, pp. 1670–1677 (2006). IEEE

    work page 2006

  4. [4]

    IEEE transactions on cybernetics46(2), 487–498 (2015)

    Han, J., Zhang, D., Wen, S., Guo, L., Liu, T., Li, X.: Two-stage learning to predict human eye fixations via sdaes. IEEE transactions on cybernetics46(2), 487–498 (2015)

    work page 2015

  5. [5]

    IEEE Transactions on Circuits and Systems for Video Technology18(10), 1397–1410 (2008)

    Tao, D., Song, M., Li, X., Shen, J., Sun, J., Wu, X., Faloutsos, C., Maybank, S.J.: Bayesian tensor approach for 3-d face modeling. IEEE Transactions on Circuits and Systems for Video Technology18(10), 1397–1410 (2008)

    work page 2008

  6. [6]

    Science robotics8(80), 9548 (2023)

    Arm, P., Waibel, G., Preisig, J., Tuna, T., Zhou, R., Bickel, V., Ligeza, G., Miki, T., Kehl, F., Kolvenbach, H.,et al.: Scientific exploration of challenging planetary analog environments with a team of legged robots. Science robotics8(80), 9548 (2023)

    work page 2023

  7. [7]

    Science Robotics7(66), 9742 (2022)

    Tranzatto, M., Miki, T., Dharmadhikari, M., Bernreiter, L., Kulkarni, M., Mas- carich, F., Andersson, O., Khattak, S., Hutter, M., Siegwart, R.,et al.: Cerberus in the darpa subterranean challenge. Science Robotics7(66), 9742 (2022)

    work page 2022

  8. [8]

    arXiv preprint arXiv:2501.15830 (2025)

    Qu, D., Song, H., Chen, Q., Yao, Y., Ye, X., Ding, Y., Wang, Z., Gu, J., Zhao, B., Wang, D., et al.: Spatialvla: Exploring spatial representations for visual-language- action model. arXiv preprint arXiv:2501.15830 (2025)

    arXiv 2025

Show all 67 references
  1. [9]

    arXiv preprint arXiv:2505.21432 (2025)

    Song, H., Qu, D., Yao, Y., Chen, Q., Lv, Q., Tang, Y., Shi, M., Ren, G., Yao, M., Zhao, B., et al.: Hume: Introducing system-2 thinking in visual-language-action model. arXiv preprint arXiv:2505.21432 (2025)

    2025 arXiv

  2. [10]

    https://arxiv.org/abs/2506.12479 14

    An, H., Hu, W., Huang, S., Huang, S., Li, R., Liang, Y., Shao, J., Song, Y., Wang, Z., Yuan, C., Zhang, C., Zhang, H., Zhuang, W., Li, X.: AI Flow: Perspectives, Scenarios, and Approaches (2025). https://arxiv.org/abs/2506.12479 14

    2025 arXiv

  3. [11]

    arXiv preprint arXiv:2506.08840 (2025)

    Wang, D., Wang, X., Liu, X., Shi, J., Zhao, Y., Bai, C., Li, X.: More: Mixture of residual experts for humanoid lifelike gaits learning on complex terrains. arXiv preprint arXiv:2506.08840 (2025)

    2025 arXiv

  4. [12]

    IEEE Robotics and Automation Letters3(3), 1560–1567 (2018)

    Winkler, A.W., Bellicoso, C.D., Hutter, M., Buchli, J.: Gait and trajectory opti- mization for legged systems through phase-based end-effector parameterization. IEEE Robotics and Automation Letters3(3), 1560–1567 (2018)

    2018

  5. [13]

    Science robotics5(47), 5986 (2020)

    Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., Hutter, M.: Learning quadrupedal locomotion over challenging terrain. Science robotics5(47), 5986 (2020)

    2020

  6. [14]

    In: Robotics: Science and Systems (2021)

    Kumar, A., Fu, Z., Pathak, D., Malik, J.: Rma: Rapid motor adaptation for legged robots. In: Robotics: Science and Systems (2021)

    2021

  7. [15]

    Science robotics 7(62), 2822 (2022)

    Miki, T., Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., Hutter, M.: Learning robust perceptive locomotion for quadrupedal robots in the wild. Science robotics 7(62), 2822 (2022)

    2022

  8. [16]

    In: Conference on Robot Learning, pp

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

    2022

  9. [17]

    arXiv preprint arXiv:2108.10470 (2021)

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

    2021 arXiv

  10. [18]

    In: Proceedings of the 26th Annual International Conference on Machine Learning, pp

    Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th Annual International Conference on Machine Learning, pp. 41–48 (2009)

    2009

  11. [19]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Roth, P., Nubert, J., Yang, F., Mittal, M., Hutter, M.: Viplanner: Visual semantic imperative learning for local navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 5243–5249 (2024). IEEE

    2024

  12. [20]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA), pp

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

    2023

  13. [21]

    IEEE Robotics and Automation Letters7(2), 2353–2360 (2022)

    Mattamala, M., Chebrolu, N., Fallon, M.: An efficient locally reactive controller for safe navigation in visual teach and repeat missions. IEEE Robotics and Automation Letters7(2), 2353–2360 (2022)

    2022

  14. [22]

    arXiv preprint arXiv:2305.14654 (2023) 15

    Caluwaerts, K., Iscen, A., Kew, J.C., Yu, W., Zhang, T., Freeman, D., Lee, K.- H., Lee, L., Saliceti, S., Zhuang, V., et al.: Barkour: Benchmarking animal-level agility with quadruped robots. arXiv preprint arXiv:2305.14654 (2023) 15

    2023 arXiv

  15. [23]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Xu, S., Zhang, W., Ho, C.P., Zhu, L.: Optimal prescribed-time control based reac- tive planning system for quadruped robot navigation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 13185–13191 (2024). IEEE

    2024

  16. [24]

    arXiv preprint arXiv:2404.15256 (2024)

    Ren, J., Liu, Y., Dai, Y., Wang, G.: Top-nav: Legged navigation integrating ter- rain, obstacle and proprioception estimation. arXiv preprint arXiv:2404.15256 (2024)

    2024 arXiv

  17. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Fu, Z., Kumar, A., Agarwal, A., Qi, H., Malik, J., Pathak, D.: Coupling vision and proprioception for navigation of legged robots. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17273–17283 (2022)

    2022

  18. [26]

    In: 7th Annual Conference on Robot Learning (2023)

    Karnan, H., Yang, E., Farkash, D., Warnell, G., Biswas, J., Stone, P.: Ster- ling: Self-supervised terrain representation learning from unconstrained robot experience. In: 7th Annual Conference on Robot Learning (2023)

    2023

  19. [27]

    In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

    Truong, J., Yarats, D., Li, T., Meier, F., Chernova, S., Batra, D., Rai, A.: Learn- ing navigation skills for legged robots with learned robot embeddings. In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 484–491 (2021). IEEE

    2021

  20. [28]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Cheng, X., Shi, K., Agarwal, A., Pathak, D.: Extreme parkour with legged robots. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 11443–11450 (2024). IEEE

    2024

  21. [29]

    In: Conference on Robot Learning (CoRL) (2023)

    Zhuang, Z., Fu, Z., Wang, J., Atkeson, C., Schwertfeger, S., Finn, C., Zhao, H.: Robot parkour learning. In: Conference on Robot Learning (CoRL) (2023)

    2023

  22. [30]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Barasuol, V., Emre, S., Medeiros, V.S., Bratta, A., Semini, C.: Introducing the carpal-claw: a mechanism to enhance high-obstacle negotiation for quadruped robots. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 3457–3463 (2024). IEEE

    2024

  23. [31]

    In: Robotics: Science and Systems (RSS) (2024)

    He, T., Zhang, C., Xiao, W., He, G., Liu, C., Shi, G.: Agile but safe: Learning collision-free high-speed legged locomotion. In: Robotics: Science and Systems (RSS) (2024)

    2024

  24. [32]

    Science Robotics9(89), 9641 (2024)

    Lee, J., Bjelonic, M., Reske, A., Wellhausen, L., Miki, T., Hutter, M.: Learning robust autonomous navigation and locomotion for wheeled-legged robots. Science Robotics9(89), 9641 (2024)

    2024

  25. [33]

    Science Robotics9(88), 7566 (2024)

    Hoeller, D., Rudin, N., Sako, D., Hutter, M.: Anymal parkour: Learning agile navigation for quadrupedal robots. Science Robotics9(88), 7566 (2024)

    2024

  26. [34]

    IEEE Transactions on Robotics 16 39(5), 3402–3421 (2023)

    Grandia, R., Jenelten, F., Yang, S., Farshidian, F., Hutter, M.: Perceptive locomo- tion through nonlinear model-predictive control. IEEE Transactions on Robotics 16 39(5), 3402–3421 (2023)

    2023

  27. [35]

    In: Conference on Robot Learning, pp

    Margolis, G.B., Agrawal, P.: Walk these ways: Tuning robot control for gener- alization with multiplicity of behavior. In: Conference on Robot Learning, pp. 22–31 (2023). PMLR

    2023

  28. [36]

    The International Journal of Robotics Research43(4), 572–587 (2024)

    Margolis, G.B., Yang, G., Paigwar, K., Chen, T., Agrawal, P.: Rapid locomotion via reinforcement learning. The International Journal of Robotics Research43(4), 572–587 (2024)

    2024

  29. [37]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Shi, J., Bai, C., He, H., Han, L., Wang, D., Zhao, B., Zhao, M., Li, X., Li, X.: Robust quadrupedal locomotion via risk-averse policy learning. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 11459–11466 (2024). IEEE

    2024

  30. [38]

    In: The Twelfth International Conference on Learning Representations (2024)

    Long, J., Wang, Z., Li, Q., Cao, L., Gao, J., Pang, J.: Hybrid internal model: Learning agile legged locomotion with simulated robot response. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  31. [39]

    arXiv preprint arXiv:2404.18225 (2024)

    Cheng, Y., Liu, H., Pan, G., Ye, L., Liu, H., Liang, B.: Quadruped robot traversing 3d complex environments with limited perception. arXiv preprint arXiv:2404.18225 (2024)

    2024 arXiv

  32. [40]

    arXiv preprint arXiv:2405.12424 (2024)

    Shi, F., Zhang, C., Miki, T., Lee, J., Hutter, M., Coros, S.: Rethinking robust- ness assessment: Adversarial attacks on learning-based quadrupedal locomotion controllers. arXiv preprint arXiv:2405.12424 (2024)

    2024 arXiv

  33. [41]

    In: International Conference on Learning Representations (2022)

    Yang, R., Zhang, M., Hansen, N., Xu, H., Wang, X.: Learning vision- guided quadrupedal locomotion end-to-end with cross-modal transform- ers. In: International Conference on Learning Representations (2022). https://openreview.net/forum?id=nhnJ3oo6AB

    2022

  34. [42]

    arXiv preprint arXiv:2403.00187 (2024)

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

    2024 arXiv

  35. [43]

    In: Conference on Robot Learning, pp

    Agarwal, A., Kumar, A., Malik, J., Pathak, D.: Legged locomotion in challenging terrains using egocentric vision. In: Conference on Robot Learning, pp. 403–415 (2023). PMLR

    2023

  36. [44]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA), pp

    Seo, M., Gupta, R., Zhu, Y., Skoutnev, A., Sentis, L., Zhu, Y.: Learning to walk by steering: Perceptive quadrupedal locomotion in dynamic environments. In: 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 5099–5105 (2023). IEEE

    2023

  37. [45]

    arXiv preprint arXiv:2309.15462 (2023) 17

    Jenelten, F., He, J., Farshidian, F., Hutter, M.: Dtc: Deep tracking control–a uni- fying approach to model-based planning and reinforcement-learning for versatile and robust locomotion. arXiv preprint arXiv:2309.15462 (2023) 17

    2023 arXiv

  38. [46]

    IEEE Robotics and Automation Letters (2023)

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

    2023

  39. [47]

    In: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

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

    2022

  40. [48]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Zhang, C., Jin, J., Frey, J., Rudin, N., Mattamala, M., Cadena, C., Hutter, M.: Resilient legged local navigation: Learning to traverse with compromised per- ception end-to-end. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 34–41 (2024). IEEE

    2024

  41. [49]

    arXiv preprint arXiv:2311.10484 (2023)

    Zhang, C., Rudin, N., Hoeller, D., Hutter, M.: Learning agile locomotion on risky terrains. arXiv preprint arXiv:2311.10484 (2023)

    2023 arXiv

  42. [50]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp

    Cheng, J., Vlastelica, M., Kolev, P., Li, C., Martius, G.: Learning diverse skills for local navigation under multi-constraint optimality. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 5083–5089 (2024). IEEE

    2024

  43. [51]

    In: 2018 IEEE International Conference on Robotics and Automation (ICRA), pp

    Faust, A., Oslund, K., Ramirez, O., Francis, A., Tapia, L., Fiser, M., David- son, J.: Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling-based planning. In: 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 511...

    2018

  44. [52]

    Autonomous Robots46(5), 569–597 (2022)

    Xiao, X., Liu, B., Warnell, G., Stone, P.: Motion planning and control for mobile robot navigation using machine learning: a survey. Autonomous Robots46(5), 569–597 (2022)

    2022

  45. [53]

    In: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp

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

    2021

  46. [54]

    In: 2017 Ieee International Conference on Robotics and Automation (icra), pp

    Pfeiffer, M., Schaeuble, M., Nieto, J., Siegwart, R., Cadena, C.: From percep- tion to decision: A data-driven approach to end-to-end motion planning for autonomous ground robots. In: 2017 Ieee International Conference on Robotics and Automation (icra), pp. 1527–1533 (2017). IEEE

    2017

  47. [55]

    In: 2017 IEEE International Conference on Robotics and Automation (ICRA), pp

    Zhu, Y., Mottaghi, R., Kolve, E., Lim, J.J., Gupta, A., Fei-Fei, L., Farhadi, A.: Target-driven visual navigation in indoor scenes using deep reinforcement learning. In: 2017 IEEE International Conference on Robotics and Automation (ICRA), pp. 3357–3364 (2017). IEEE

    2017

  48. [56]

    In: 7th Annual Conference 18 on Robot Learning (2023).https://arxiv.org/abs/2306.14846

    Shah, D., Sridhar, A., Dashora, N., Stachowicz, K., Black, K., Hirose, N., Levine, S.: ViNT: A foundation model for visual navigation. In: 7th Annual Conference 18 on Robot Learning (2023).https://arxiv.org/abs/2306.14846

    2023 arXiv

  49. [57]

    arXiv preprint arXiv:2407.03122 (2024)

    Gao, W., Ai, B., Loo, J., Hsu, D., et al.: Intentionnet: Map-lite visual navigation at the kilometre scale. arXiv preprint arXiv:2407.03122 (2024)

    2024 arXiv

  50. [58]

    arXiv preprint arXiv:2406.16578 (2024)

    Wang, Y., Mei, Y., Zheng, S., Jin, Q.: Quadrupedgpt: Towards a versatile quadruped agent in open-ended worlds. arXiv preprint arXiv:2406.16578 (2024)

    2024 arXiv

  51. [59]

    arXiv preprint arXiv:2407.02666 (2024)

    Chen, A.S., Lessing, A.M., Tang, A., Chada, G., Smith, L., Levine, S., Finn, C.: Commonsense reasoning for legged robot adaptation with vision-language models. arXiv preprint arXiv:2407.02666 (2024)

    2024 arXiv

  52. [60]

    arXiv preprint arXiv:1707.06347 (2017)

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

    2017 arXiv

  53. [61]

    In: arXiv Preprint arXiv:2303.03378 (2023)

    Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., Huang, W., Chebotar, Y., Sermanet, P., Duckworth, D., Levine, S., Vanhoucke, V., Hausman, K., Toussaint, M., Greff, K., Zeng, A., Mordatch, I., Florence, P.: ...

    2023 arXiv

  54. [62]

    IEEE Transactions on Neural Networks and Learning Systems35(6), 8708–8714 (2022)

    Li, X.: Positive-incentive noise. IEEE Transactions on Neural Networks and Learning Systems35(6), 8708–8714 (2022)

    2022

  55. [63]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Zhang, H., Huang, S., Guo, Y., Li, X.: Variational positive-incentive noise: How noise benefits models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  56. [64]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Huang, S., Zhang, H., Li, X.: Enhance vision-language alignment with noise. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 17449–17457 (2025)

    2025

  57. [65]

    arXiv preprint arXiv:2408.09929 (2024)

    Zhang, H., Xu, Y., Huang, S., Li, X.: Data augmentation of contrastive learning is estimating positive-incentive noise. arXiv preprint arXiv:2408.09929 (2024)

    2024 arXiv

  58. [66]

    arXiv preprint arXiv:2505.19024 (2025)

    Huang, S., Xu, Y., Zhang, H., Li, X.: Learn beneficial noise as graph augmenta- tion. arXiv preprint arXiv:2505.19024 (2025)

    2025 arXiv

  59. [67]

    arXiv preprint arXiv:2303.08774 (2023) 19

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023) 19

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.