Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Attention-Based Map Encoding for Learning Generalized Legged Locomotion

T0 review · 5 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 An attention-based map encoder lets an end-to-end learned controller place feet precisely on sparse terrain and reveals which future foothold it has chosen.

desk verdict Solid architecture paper with a genuinely new map-encoding idea; the hardware stories are compelling, but the GR-1 robustness claim is undercut by motion-capture ground truth and the lack of quantitative metrics. read the letter →

arxiv 2506.09588 v1 pith:SIOV6QBH submitted 2025-06-11 cs.RO

Junzhe He , Chong Zhang , Fabian Jenelten , Ruben Grandia , Moritz BÄcher , Marco Hutter This is my paper · ORCID
classification cs.RO
keywords leggedlocomotionreinforcementlearningattentionmechanismmulti-headterrainperceptionsim-to-realtransfersparsehumanoidrobot
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 attempts to close a gap in legged locomotion: learning-based controllers are robust but imprecise on terrain where safe footholds are scarce, while model-based planners are precise but fragile under real-world uncertainty. The authors propose an end-to-end controller trained with reinforcement learning in which a convolutional network reads a robot-centric height map and a multi-head attention module, using the robot's own state as the query, condenses that map into a compact encoding. Across simulation and real-world runs on an ANYmal-D quadruped and a GR-1 humanoid, the controller places feet precisely on stepping stones, beams, and gaps, and it generalizes to terrain layouts never seen in training. The paper further claims the attention weights themselves point at the next feasible foothold without any supervised label, offering a way to see what the network perceives.

What carries the argument

The load-bearing component is the attention-based map encoding: a CNN with kernel size 5 extracts per-point local features from an L by W robot-centric height map, the per-point features are concatenated with their three-dimensional coordinates, and a multi-head attention module with 16 heads takes a linear embedding of proprioception as the query and the terrain features as keys and values. The query-conditioned weighted sum produces a compact map encoding that the policy network turns into joint actions, so the representation is state-dependent rather than a fixed feature vector; this is what lets the controller direct its focus to future footholds and, in the paper's view, gives it both precision and generalization.

What would settle it

Deploy the controller on a fixed sparse-terrain course, for example twenty trials across a 12-cm stepping-stone field, with matched velocity commands and count traversals; if real-world success is far below the roughly 100 percent the same controller achieves on the simulated parkour, the zero-shot transfer claim is weakened. Alternatively, add an artificial bias to the base-velocity estimates during deployment and check whether the attention weights move off steppable regions; if they do not, the proprioceptive conditioning is not actually driving foothold selection.

Watch

Extended reading notes

Core claim

The paper's central claim is that a proprioception-conditioned attention mechanism over learned point-wise terrain features can serve as the map encoder of an end-to-end reinforcement-learning locomotion controller, and that this single architecture yields precise foothold selection on sparse terrain, robustness to disturbances, and generalization across terrain types and embodiments. The authors report that the attention module learns, without supervised labels, to assign the highest weights to map cells that are steppable and reachable as future footholds, that a two-stage training schedule (first privileged terrain, then noisy perception with added terrain variety) is necessary for this to emerge, and that the resulting controller outperforms the DTC hybrid and a prior RL baseline in success rate and velocity tracking in simulation while transferring zero-shot to hardware on both a 12-DoF quadruped and a 23-DoF humanoid.

Load-bearing premise

Zero-shot transfer from simulation to real hardware is the load-bearing premise: the simulated terrains, observation noise and drift, and domain randomization ranges must faithfully reproduce real elevation-mapping error, foot slippage, and actuator behavior, yet the paper reports no quantitative real-world success or failure rates to confirm it.

Editorial extensions

If this is right

  • End-to-end learning can replace the model-based planner in sparse-terrain locomotion: the controller attains a 100 percent success rate on an unseen obstacle parkour in simulation and does not need a model predictive controller to run at deployment.
  • The same network topology, with only the map-scan dimensions changed, transfers from a quadruped to a humanoid controller, implying the attention-based map encoding generalizes across embodiments and kinematics.
  • Attention weights give a built-in interpretability channel: engineers can visualize which terrain regions the policy uses for the next step, which can guide debugging and safety analysis.
  • Because the policy can adjust gait frequency and recover from slippage using whole-body motions, the controller exhibits agility, including knee and arm use and in-air foot switching, that model-based contact state machines do not readily produce.
  • The two-stage training schedule is a prerequisite: training from scratch on all terrains or with noise from the start converges to worse terrain levels, so staged terrain and perception curricula are part of the method's result.

Reading between the lines

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

  • Beyond the paper: the attention weights could be exported as a learned foothold-cost map and used to warm-start or constrain a model-based planner, potentially combining the controller's perception with planning guarantees.
  • A testable extension is to measure whether the attention module's selected footholds remain consistent when proprioceptive inputs are artificially biased; if they shift incorrectly, the proprioceptive conditioning is less reliable than claimed.
  • Since the method uses a 2.5-D height map, the same architecture could be extended to 3-D occupancy or signed-distance representations, a direction the paper's own limitation discussion anticipates.
  • The reported real-world results are qualitative demonstrations; a quantitative benchmark with repeated trials and failure counts across terrain difficulties would let others compare the zero-shot transfer claims directly.
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

5 major / 5 minor

Summary. The paper proposes an attention-based map encoding for end-to-end reinforcement-learned legged locomotion. A CNN extracts local features from a robot-centric height map, and a multi-head attention module uses the proprioceptive embedding as a query to produce a compact map encoding. The controller is trained in two stages: first on base terrains with privileged observations, then with additional terrains and injected perception noise/drift. The method is evaluated on two embodiments, ANYmal-D (quadruped) and Fourier GR-1 (humanoid), in simulation against DTC and a baseline RL controller, with ablations of the two-stage pipeline and network structure, and qualitatively on real hardware. The authors claim precise, robust, generalized locomotion on sparse terrains, as well as an interpretable representation in which attention weights concentrate on future footholds.

Significance. If the claims are substantiated, the paper would be a meaningful advance: it demonstrates that an end-to-end DRL controller with a map-attention encoder can handle sparse terrain that typically requires model-based planning, and it does so on two different embodiments. The paper has notable strengths: a clear architectural motivation, a sensible two-stage training scheme, ablations isolating both the training pipeline and the attention mechanism, public data for reproducing the plots, and real-hardware demonstrations on both a quadruped and a humanoid. However, several load-bearing claims currently rest on qualitative demonstrations or single-seed simulation results, and the real-world evidence for robustness to perception uncertainty is incomplete.

major comments (5)
  1. [Deployment (p.15)] The real GR-1 experiments do not exercise the perception-robustness claim. The text states that for GR-1, "the map scans are sampled through ray-casting on a pre-designed terrain mesh given the robot's pose captured by the Qualisys Motion Capture system." This means the humanoid's hardware deployment used ground-truth terrain geometry and external localization, bypassing the elevation-mapping noise and drift that the method is claimed to be robust against. The central claim of robust sim-to-real transfer for GR-1 is therefore unsupported by the presented real-world evidence.
  2. [Section D, Simulation-based Evaluations (Fig. 6)] All simulation benchmarks are single-seed point estimates. The text says, "All controllers are deployed in the same simulated environment with observation noises and drifts sampled with the same random seed." Without multiple seeds or error bars, the reported differences (e.g., "26.5% and 77.3% higher success rates" in Fig. 6A b, and the ablation comparisons in Fig. 6B/C) cannot be distinguished from stochastic variation. Additionally, the velocity tracking errors in Fig. 6A a are computed only for surviving agents, which is a selection bias if survival rates differ across controllers.
  3. [Results A and Deployment (Fig. 3)] Real-world results are presented only qualitatively. No trial counts, success/failure/stuck rates, or quantitative precision metrics (e.g., foot placement errors relative to commanded footholds) are reported for ANYmal-D or GR-1. The text claims the controllers "achieve high robustness in real-world deployments" and "precise and generalized performance," but these assertions are not backed by numbers. This evidence gap is load-bearing because the headline contribution is sim-to-real transfer on real sparse terrains.
  4. [Results E, Interpretable Attention-based Map Encoding (Fig. 7)] The claim that attention weights "indicate the next foothold" or "are concentrated around the next steppable region" is supported only by visual inspection of a few examples. The paper should provide a quantitative evaluation, such as comparing the locations of high attention weights with the actual subsequent contact positions, or measuring the correlation between attention maps and feasible foothold areas. As presented, the interpretability claim remains subjective and could be a post-hoc rationalization.
  5. [Domain Randomization (p.13-14)] The magnitudes of the injected sensor noise, map drift, artificial pushes, and randomization ranges for mass and friction are not specified. These parameters define the uncertainty against which the method is claimed to be robust, and stage 2 is explicitly designed to close the sim-to-real gap. Without these values, the experiments cannot be reproduced and the robustness claim cannot be assessed quantitatively.
minor comments (5)
  1. [Table 2] The layout of weight and index columns for ANYmal-D and GR-1 is difficult to parse, especially where some reward terms have entries for both robots and others only for one. A clearer tabular format or a separate table per robot would improve readability.
  2. [Figure 6] The bar charts in Figure 6 would be substantially more informative with error bars or at least a statement of the number of seeds and episodes used for each success-rate estimate.
  3. [Deployment (p.15)] The motion-capture system is spelled "Qualysis" in the text but correctly as "Qualisys" in reference 43; please unify the spelling.
  4. [Figure 7C caption] The caption refers to "fine-tuning terrains" while the body text describes these as "unseen terrains" evaluated with the stage 1 controller. Please clarify whether these terrains were seen in stage 2 and whether the visualization is from the stage 1 or stage 2 controller.
  5. [Observation Space (p.13)] The map scan description says "a vector map scans" with L × W × 3 points, but the three channels are not explicitly defined in the text (they appear to be x, y, z coordinates in the robot frame). Please state this explicitly and specify the map resolution and extents.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical RL results, and the attention-foothold observation is an emergent post-hoc finding rather than a fitted target.

full rationale

This paper contains no formal derivation chain whose output is equivalent to its input. The load-bearing claims are empirical: an RL-trained CNN plus multi-head attention policy achieves high success rates on held-out sparse terrains in simulation and qualitative hardware demos. The attention-weight visualizations are post-hoc observations of a trained policy, not labels fitted into the network; the paper explicitly states the foothold indication arises 'without any supervised learning,' and the weights are free variables optimized only for the locomotion reward. The main benchmark comparisons are to DTC [1] and baseline-rl [13], both prior works with overlapping authorship, but the reported numbers are simulation evaluations against published methods, not parameters fitted to the target, and the paper's central architecture and two-stage training pipeline do not reduce to those citations. The two-stage training pipeline is an experimental design choice, and the 'prediction' of generalization to unseen terrains is a held-out evaluation rather than a quantity constructed from the training fit. The lack of quantitative real-world metrics, and the use of motion-capture ground-truth maps for GR-1, weaken the evidential support for the sim-to-real robustness claim but do not constitute circularity. Therefore no circular step meeting the quoted-reduction standard was found.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on many hand-chosen design choices, including architecture dimensions, reward weights, terrain curriculum, and domain randomization ranges, plus the standard sim-to-real assumption that simulation fidelity is adequate. No new physical entities are introduced; the attention module is a neural network component from prior work, not a new postulated entity.

free parameters (5)
  • MHA dimension d and number of heads h = d=64, h=16
    Chosen by hand; no sensitivity study shown.
  • Map scan dimensions and resolution = 26x16 (ANYmal-D), 17x11 (GR-1), 10 cm resolution
    Heuristically chosen; assumed sufficient for sparse terrains.
  • Reward weights in Table 2 = Multiple per-term weights, see Table 2
    Tuned by hand; no ablation over reward weights is reported.
  • Domain randomization ranges = Not stated explicitly
    Noise, drift, push, mass, and friction ranges are described qualitatively but exact values are omitted, making reproducibility harder.
  • Stage-wise training durations = ANYmal-D 18000+3600 epochs; GR-1 15000+3200 epochs
    Chosen empirically; no analysis of sensitivity to these durations.
assumptions (3)
  • domain assumption Simulation with domain randomization transfers zero-shot to real hardware.
    Stated in Domain Randomization and Deployment; no quantitative hardware metrics validate this.
  • domain assumption The terrain is adequately represented by a 2.5D height map at 10 cm resolution.
    This is an acknowledged limitation in the Discussion; the paper assumes it holds for all tested sparse terrains.
  • domain assumption Reward weights and curriculum define a good proxy for robust locomotion and foothold selection.
    The learned behavior and emergent attention depend on the hand-designed reward and curriculum; no proof that these incentives are the right ones.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Attention-Based Map Encoding for Learning Generalized Legged Locomotion." pith.science (2026). https://pith.science/paper/SIOV6QBH

@misc{pith2026250609588,
  author       = {Pith},
  title        = {Pith review of: Attention-Based Map Encoding for Learning Generalized Legged Locomotion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIOV6QBH}},
  note         = {Machine review of arXiv:2506.09588}
}
read the original abstract

Dynamic locomotion of legged robots is a critical yet challenging topic in expanding the operational range of mobile robots. It requires precise planning when possible footholds are sparse, robustness against uncertainties and disturbances, and generalizability across diverse terrains. While traditional model-based controllers excel at planning on complex terrains, they struggle with real-world uncertainties. Learning-based controllers offer robustness to such uncertainties but often lack precision on terrains with sparse steppable areas. Hybrid methods achieve enhanced robustness on sparse terrains by combining both methods but are computationally demanding and constrained by the inherent limitations of model-based planners. To achieve generalized legged locomotion on diverse terrains while preserving the robustness of learning-based controllers, this paper proposes to learn an attention-based map encoding conditioned on robot proprioception, which is trained as part of the end-to-end controller using reinforcement learning. We show that the network learns to focus on steppable areas for future footholds when the robot dynamically navigates diverse and challenging terrains. We synthesize behaviors that exhibit robustness against uncertainties while enabling precise and agile traversal of sparse terrains. Additionally, our method offers a way to interpret the topographical perception of a neural network. We have trained two controllers for a 12-DoF quadrupedal robot and a 23-DoF humanoid robot respectively and tested the resulting controllers in the real world under various challenging indoor and outdoor scenarios, including ones unseen during training.

Figures

Figures reproduced from arXiv: 2506.09588 by the authors.

Figure 1
Figure 1. Learning interpretable, generalizable, agile, and robust legged locomotion on diverse terrains. Our end-to-end controllers enabled ANYmal-D (a - i) and GR-1 (j - o) to dynamically traverse diverse challenging terrains. Highly interpretable point-wise map encodings are graphically visualized in k, m, and o, where colors more intense in red represent higher attention weights, indicating the next foothold [PITH_FULL_I… view at source ↗
Figure 2
Figure 2. Precise and Generalized Locomotion in Simulation. (A) Selection of base terrains for GR1 stage 1 training, including grid stones (a), pallets (b), beams (c), and gaps (d) (B) Selection of fine-tuning terrains for GR1 stage 2 training, including pentagon stones (a), single￾column stones (b), narrow pallets (c), and consecutive gaps (d) (C) GR1 and ANYmal-D on obstacle parkour [23]. (D) Selection of fine-tuning terrai… view at source ↗
Figure 3
Figure 3. Precise and Generalized Locomotion on Hardware. (A) Selection of sparse terrains for GR1, including beam and gap (a) and single￾column stones with height difference (b). (B) Selection of sparse terrains for ANYmal-D, including stepping stones (a), stepping stones with height difference (b), randomly placed stepping stones forward (c), randomly placed stepping stones sideway (d), boxes and gaps (e), and a 19- cm wide… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Agility and Recovery Reflexes by Whole-Body Coordination. (A) ANYmal-D using its knee to climb up a large rock while rotating the torso. (B) ANYmal-D recovered from penetration of feet into the shaky debris due to slippage by knee support. (C) GR-1 traversing one row o…
Figure 5
Figure 5. Figure 5: Versatile Velocity Tracking. Our learned controllers demonstrated versatile velocity tracking capabilities on ANYmal-D and GR-1. (A) ANYmal-D maneuvering on the debris, overcoming sparse terrain with movable supports and showcasing omnidirectional versatility. (B) GR-1…
Figure 6
Figure 6. Figure 6: Simulation-based Evaluations. We only evaluated the performance of our approach and benchmarked it with other methods on ANYmal-D. (A) Benchmark with DTC and Baseline-rl. (a) Our method shows overall lower velocity tracking errors for different forward velocity command…
Figure 7
Figure 7. Figure 7: Attention Weights Visualization. We visualized the height scans with associated attention weights on each scan, where colors with higher intensity in red indicate higher attention. (A) Attention weights on a mixture of terrains for a stage 2 controller. (B) Attention w…
Figure 8
Figure 8. Figure 8: Proposed Control Pipeline. (A) A locomotion strategies abstract. (B) Stage 1 training.(C) Stage 2 fine-tuning with disturbance, uncer￾tainty, and extra harder terrains [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Humanoid Occupancy: Enabling A Generalized Multimodal Occupancy Perception System on Humanoid Robots

    cs.RO 2025-07 conditional novelty 5.0 of 10

    A humanoid-specific multimodal occupancy perception system with a new dataset, sensor layout, and a fusion network that claims state-of-the-art results on its own benchmark.

Reference graph

Works this paper leans on

44 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    Jenelten, J

    F . Jenelten, J. He, F . Farshidian, M. Hutter, Dtc: Deep tracking control, Science Roboticsp. eadh5401 (2024)

    work page 2024

  2. [2]

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

    2020

  3. [3]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hutter, Learn- ing robust perceptive locomotion for quadrupedal robots in the wild, Science Roboticsp. eabk2822 (2022)

    2022

  4. [4]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Roboticsp. eaau5872 (2019)

    work page 2019

  5. [5]

    Siekmann, K

    J. Siekmann, K. Green, J. Warila, A. Fern, J. W. Hurst, Blind bipedal stair traversal via sim-to-real reinforcement learning,Robotics: Sci- ence and Systems XVII, Virtual Event, July 12-16, 2021, D. A. Shell, M. Toussaint, M. A. Hsieh, eds. (2021)

    work page 2021

  6. [6]

    Rudin, D

    N. Rudin, D. Hoeller, P . Reist, M. Hutter, Learning to walk in minutes using massively parallel deep reinforcement learning,Proceedings of the 5th Conference on Robot Learning, A. Faust, D. Hsu, G. Neumann, eds., 91–100 (PMLR, 2022)

    work page 2022

  7. [7]

    Rudin, D

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

    work page 2022

  8. [8]

    Hoeller, N

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

    work page 2024

Show all 44 references
  1. [9]

    Zhuang, Z

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

    2023

  2. [10]

    Cheng, K

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

    2024

  3. [11]

    Zhuang, S

    Z. Zhuang, S. Y ao, H. Zhao, Humanoid parkour learning,8th Annual Conference on Robot Learning(2024)

    2024

  4. [12]

    Zhang, W

    C. Zhang, W. Xiao, T. He, G. Shi, Wococo: Learning whole-body humanoid control with sequential contacts,8th Annual Conference on Robot Learning(2024)

    2024

  5. [13]

    Zhang, N

    C. Zhang, N. Rudin, D. Hoeller, M. Hutter, Learning agile locomotion on risky terrains,2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(2024)

    2024

  6. [14]

    H. Duan, A. Malik, M. S. Gadde, J. Dao, A. Fern, J. W. Hurst, Learning dynamic bipedal walking across stepping stones,2022 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS)6746–6752 (2022)

    2022

  7. [15]

    Shkolnik, M

    A. Shkolnik, M. Levashov, I. R. Manchester, R. Tedrake, Bounding on rough terrain with the littledog robot,The International Journal of Robotics Research192–215 (2011)

    2011

  8. [16]

    Fankhauser, M

    P . Fankhauser, M. Bjelonic, C. Dario Bellicoso, T. Miki, M. Hutter, Robust rough-terrain locomotion with a quadrupedal robot,2018 IEEE International Conference on Robotics and Automation (ICRA), 5761– 5768 (2018)

    2018

  9. [17]

    Dimitrov, A

    D. Dimitrov, A. Sherikov, P .-B. Wieber, A sparse model predictive control formulation for walking motion generation,2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2292–2299 (2011)

    2011

  10. [18]

    Neunert, C

    M. Neunert, C. de Crousaz, F . Furrer, M. Kamel, F . Farshidian, R. Y . Siegwart, J. Buchli, Fast nonlinear model predictive control for unified trajectory optimization and tracking,2016 IEEE International Conference on Robotics and Automation (ICRA)1398–1404 (2016)

    2016

  11. [19]

    Farshidian, E

    F . Farshidian, E. Jelavic, A. Satapathy, M. Giftthaler, J. Buchli, Real- time motion planning of legged robots: A model predictive control approach,2017 IEEE-RAS 17th International Conference on Humanoid Robotics (Humanoids)577–584 (2017)

    2017

  12. [20]

    Di Carlo, P

    J. Di Carlo, P . M. Wensing, B. Katz, G. Bledt, S. Kim, Dynamic locomo- tion in the mit cheetah 3 through convex model-predictive control,2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 1–9 (2018)

    2018

  13. [21]

    Grandia, F

    R. Grandia, F . Farshidian, R. Ranftl, M. Hutter, Feedback MPC for torque-controlled legged robots,2019 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems, IROS 2019, Macau, SAR, China, November 3-8, 2019, 4730–4737 (IEEE, 2019)

    2019

  14. [22]

    Mastalli, W

    C. Mastalli, W. Merkt, G. Xin, J. Shim, M. Mistry, I. Havoutis, S. Vijayaku- mar, Agile maneuvers in legged robots: a predictive control approach (2022)

    2022

  15. [23]

    Grandia, F

    R. Grandia, F . Jenelten, S. Y ang, F . Farshidian, M. Hutter, Perceptive locomotion through nonlinear model-predictive control,IEEE Transac- tions on Robotics1–20 (2023)

    2023

  16. [24]

    Jenelten, R

    F . Jenelten, R. Grandia, F . Farshidian, M. Hutter, Tamols: Terrain-aware motion optimization for legged systems,IEEE Transactions on Robotics 3395–3413 (2022)

    2022

  17. [25]

    T. Kwon, Y . Lee, M. Van De Panne, Fast and flexible multilegged locomotion using learned centroidal dynamics,ACM Transactions on Graphics39(2020)

    2020

  18. [26]

    Melon, M

    O. Melon, M. Geisert, D. Surovik, I. Havoutis, M. Fallon, Reliable tra- jectories for dynamic quadrupeds using analytical costs and learned initializations,2020 IEEE International Conference on Robotics and Au- tomation (ICRA), 1410–1416 (2020)

    2020

  19. [27]

    Surovik, O

    D. Surovik, O. Melon, M. Geisert, M. Fallon, I. Havoutis, Learning an expert skill-space for replanning dynamic quadruped locomotion over obstacles,Proceedings of the 2020 Conference on Robot Learning, J. Kober, F . Ramos, C. Tomlin, eds., 1509–1518 (PMLR, 2021)

    2020

  20. [28]

    Melon, R

    O. Melon, R. Orsolino, D. Surovik, M. Geisert, I. Havoutis, M. Fal- lon, Receding-horizon perceptive trajectory optimization for dynamic legged locomotion with learned initialization,2021 IEEE International Conference on Robotics and Automation (ICRA), 9805–9811 (2021)

    2021

  21. [29]

    Gangapurwala, M

    S. Gangapurwala, M. Geisert, R. Orsolino, M. Fallon, I. Havoutis, Rloc: Terrain-aware legged locomotion using reinforcement learning and optimal control,IEEE Transactions on Robotics2908–2927 (2022)

    2022

  22. [30]

    Z. Xie, X. Da, B. Babich, A. Garg, M. v. de Panne, Glide: Generaliz- able quadrupedal locomotion in diverse environments with a centroidal model,Algorithmic F oundations of Robotics XV, S. M. LaValle, J. M. O’Kane, M. Otte, D. Sadigh, P . Tokekar, eds., 523–539 (Springer Inter...

    2023

  23. [31]

    T. Miki, L. Wellhausen, R. Grandia, F . Jenelten, T. Homberger, M. Hutter, Elevation mapping for locomotion and navigation using gpu,2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2273–2280 (2022)

    2022

  24. [32]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need,Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, R. Garnett, eds., 30(Cu...

    2017

  25. [33]

    Y ang, M

    R. Y ang, M. Zhang, N. Hansen, H. Xu, X. Wang, Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers, International Conference on Learning Representations(2022)

    2022

  26. [34]

    Radosavovic, T

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, K. Sreenath, Real-world humanoid locomotion with reinforcement learning,Science Roboticsp. eadi9579 (2024)

    2024

  27. [35]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V. Tsou- nis, J. Hwangbo, K. Bodie, P . Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, M. Hoepflinger, Anymal - a highly mobile and dynamic quadrupedal robot,2016 IEEE/RSJ International Conference on Inte...

    2016

  28. [36]

    Research Article ETH Zurich 16

    Fourier Intelligence, Fourier GR1 Robot, https://www.fftai.com/ products-gr1 (2023). Research Article ETH Zurich 16

    2023

  29. [37]

    Chignoli, D

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

    2020

  30. [38]

    Fahmi, V

    S. Fahmi, V. Barasuol, D. Esteban, O. Villarreal, C. Semini, Vital: Vision- based terrain-aware locomotion for legged robots,IEEE Transactions on Robotics885–904 (2022)

    2022

  31. [39]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Un- terthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recognition at scale,9th International Conference on Learning Rep...

    2021

  32. [40]

    T. Miki, J. Lee, L. Wellhausen, M. Hutter, Learning to walk in confined spaces using 3d representation,2024 IEEE International Conference on Robotics and Automation (ICRA), 8649–8656 (2024)

    2024

  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 Robotics and Automation Letters3699–3706 (2020)

    2020

  34. [42]

    Schulman, F

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

    2017

  35. [43]

    Qualisys, Qualisys Motion Capture system, https://www.qualisys.com/

  36. [44]

    attention-based map encoding for learning generalized legged locomotion

    J. He, C. Zhang, F . Jenelten, R. Grandia, M. Bächer, M. Hutter, Data for "attention-based map encoding for learning generalized legged locomotion", https://zenodo.org/records/14499786 (2024). ACKNOWLEDGMENTS We thank Nikita Rudin and Vladlen Koltun for their helpful discussio...

    2024

Pith tools

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