Pith. sign in

REVIEW 5 major objections 4 minor 34 references

UMC: Unified Resilient Controller for Legged Robots with Joint Malfunctions

T0 review · 5 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A single trained policy with a masking mechanism lets legged robots keep walking after sensor or joint damage, improving task completion by an average of 36% (transformer) and 39% (MLP) across three locomotion tasks.

desk verdict A useful damage taxonomy and a two-stage training recipe that clearly improve damaged-condition locomotion across four robots, but the paper's central claim that masking drives the gains is not isolated by any ablation. read the letter →

arxiv 2502.03035 v1 pith:6WDPNJ7V submitted 2025-02-05 cs.RO

Yu Qiu , Xin Lin , Jingbo Wang , Xiangtai Li , Lu Qi , Ming-Hsuan Yang This is my paper · ORCID
classification cs.RO
keywords leggedrobotsresilientlocomotionjointmalfunctionsensorfailuremaskingmechanismtwo-stagereinforcementlearningtransformerpolicyMLP
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 tries to establish that one reinforcement-learned policy can keep a legged robot functional across eight distinct damage conditions: sensor loss, restricted joint range of motion, reduced motor force, and limited joint velocity, with sensors either damaged or intact. Existing multi-policy and meta-learning approaches require switching policies or heavy maintenance, whereas UMC trains a single controller over two stages, first on healthy environments, then on damaged ones, and uses a masking mechanism during the second stage to stop the network from relying on malfunctioning joints. The authors report that this raises task-completion capability by an average of 36% for the transformer version and 39% for the MLP version across the A1, H1, and G1 locomotion tasks, and beats a multi-task fault-tolerant baseline on Solo8 by 26.8%. If the claim holds, the general lesson is that explicit masking of broken sensor and joint channels during fine-tuning can convert a normally trained locomotion policy into a damage-resilient one without per-fault policies.

What carries the argument

The load-bearing object is the mask encoder with its two-part masking strategy. The damage detection module Γ(O) outputs V (damaged joint observations zeroed) and M (an attention mask with −∞ at damaged positions), plus the three-dimensional malfunction flag F. In the transformer, M is added inside softmax attention so damaged joints contribute negligibly; in the MLP, only the zeroing step is used. Two-stage training is the other half of the machinery: pretraining on healthy environments preserves baseline mobility, and fine-tuning on uniformly sampled damage subcategories teaches the network to exploit only functional limbs. The mask strategy is what lets a single policy adapt without switching.

What would settle it

In the released code, run the trained UMC policy on the A1 task while corrupting the detection output for one joint per episode, alternately masking a healthy joint or failing to mask a damaged one, and record the 1-unit to 5-unit success rates. If these rates fall back to the normally trained baseline or below, the claimed robustness is carried by the detection assumption rather than by the policy and masking mechanism alone.

Watch

Extended reading notes

Core claim

The central claim is that a masking mechanism, applied in a two-stage training pipeline, is enough to make a single policy resilient to a broad family of sensor and joint malfunctions. The damage detection module Γ(O) produces a masked observation V that zeroes out readings from damaged joints and a masking matrix M that pushes attention weights for those joints toward zero, while a flag F toggles when joint damage is detected. The base structure — tokenizer, mask encoder, and detokenizer for the transformer, or plain hidden layers for the MLP — then computes actions using only healthy-joint information. Stage I trains normally to preserve baseline mobility; Stage II fine-tunes across normal, sensor-only, detectable joint damage, and undetectable joint damage subcategories. The paper reports average task-completion gains of 36% for the transformer and 39% for the MLP across three tasks, failure-rate reductions of 30% and 37%, and a 26.8% task-completion improvement over the MT-FTC baseline on Solo8.

Load-bearing premise

The controller's resilience relies on the damage detection module always giving a correct and immediate binary signal about which joints are damaged; if it ever masks a healthy joint or misses a damaged one, the zeroed observations and attention mask actively push the policy toward wrong actions.

Editorial extensions

If this is right

  • One policy, not a bank of fault-specific policies, covers all eight damage scenarios, so deployment and maintenance are simpler.
  • Both transformer and MLP actor networks benefit, so the masking mechanism is architecture-agnostic.
  • Normal-condition performance is retained, not sacrificed, because Stage I pretraining anchors the policy before damage fine-tuning.
  • Fall rates drop substantially on humanoid tasks, where falls are physically and financially costly.
  • The approach is model-free and needs no prior knowledge of the specific malfunction at inference time beyond the detection signal.

Reading between the lines

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

  • Because the policy is never told which joints are damaged except through the masking binary, real-world robustness is bounded by the accuracy and latency of the damage detector; the paper tests only the idealized case where detection is correct and instantaneous.
  • The same masking recipe could plausibly transfer to other robot morphologies or to richer failure modes such as partial sensor noise, since the mechanism acts on per-joint channels rather than task-specific features.
  • An ablation that corrupted the mask (false positive or false negative on one joint) would reveal how much of the reported resilience lives in the policy versus in the detection assumption; the paper does not report this experiment.
  • The gains on undetectable damage suggest that mere exposure to varied weak-joint dynamics in Stage II contributes part of the improvement, so a simpler observation-dropout baseline might capture some of the effect even without attention masking.
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 / 4 minor

Summary. The paper proposes UMC, a two-stage training framework with a masking mechanism for legged robot control under eight categories of sensor and joint damage. The method uses a damage detection module to zero out damaged-joint observations and, in the transformer variant, adds an additive attention mask; the MLP variant uses only the zeroing operation. Stage I pre-trains on normal conditions, and Stage II fine-tunes on four damage subcategories. Experiments on A1, H1, G1, and Solo8 report large improvements in distance-based task completion and reduced fall rates compared to normally trained baselines, BodyTransformer, and the MT-FTC method. Ablations cover training-stage count, sampling ratios, masking values, and training paradigm.

Significance. If the empirical claims are supported, UMC would be a useful demonstration that a single policy with a masking mechanism can ride out multiple sensor and joint failures without switching policies, which is a practically valuable property for legged robots. The paper is commendably detailed in its appendix, with explicit tables for training and inference parameters, damage ranges, and per-scenario results, and it promises public code and models. However, the central attribution of the gains to the masking mechanism is not yet isolated: no ablation removes the masking operation while preserving the two-stage damage training distribution, and for the MLP architecture no attention mask exists at all. The evaluation also relies on point estimates without error bars, and at least one appendix figure appears to be duplicated from another task, so the statistical strength of the headline numbers is not currently established.

major comments (5)
  1. [Section 4.4, Tables 6–8, Figure 4] The claim that the masking mechanism is the core contribution (Abstract; Section 3.2.1) is never isolated. None of the ablations compares two-stage training with the additive mask M in Eq. (5) against the same two-stage training without M, or removes the mask encoder while keeping the Stage-II environment distribution. This matters because Appendix B states that detectable joint damage always co-occurs with sensor damage, so the damaged-joint observations are already zeroed by the environment; the unique model component for the transformer is therefore the attention mask M, and for the MLP (§3.4) there is no attention mask at all. Please add an ablation that trains with two-stage Stage-II sampling but disables the mask (e.g., M=0), and report the task-completion metrics; without it, the 36%/39% improvement could be due to two-stage training alone.
  2. [Section 3.2.1 and Section 1 contribution list] The claim that UMC 'does not require prior knowledge of joint or limb malfunctions during inference' is inconsistent with the design, which relies on the damage detection module Γ(O) to produce a clean binary signal F and to set the correct entries of M. No detection errors, delays, or false positives are modeled in the experiments. If a real detector misses a damaged joint or masks a healthy joint, the zeroing and attention mask will actively corrupt the policy input. Please state this perfect-detection assumption explicitly, and ideally include an inference-time sensitivity analysis that perturbs the detection output.
  3. [Tables 2–5 and Figures 7–11] All quantitative claims are reported as point estimates from a single run, without error bars or multiple training seeds. Given the stochasticity of PPO training and the modest differences in some failure-rate columns (e.g., 2% vs 3% in Table 2), the significance of the reported margins is unclear. Please report mean ± std over at least three independent seeds for the main comparisons, or otherwise demonstrate that the differences are not attributable to training noise.
  4. [Section 3.3(iv) and Appendix B] The Stage-II training distribution omits scenario 5 of Table 1: undetectable ROM restriction (functional sensor with restricted range of motion). The text defines the undetectable-joint-damage subcategory as covering only reduced motor force and limited linear velocity (scenarios 6–7), yet the evaluation includes 'Undetected ROM-Limit Condition' in Figures 7–10. Please clarify whether scenario 5 is intentionally a held-out generalization test; if so, state this explicitly, since the 'eight damage scenarios' framing currently implies all are represented in the training distribution.
  5. [Appendix D, Figure 11] The figure labeled as the Solo8 SOTA comparison reproduces the same values as the G1 results in Figure 9, with the same method labels (Transformer_NM, MLP_UMC, BodyTransformer, etc.). Since Table 5's Solo8 comparison reports only aggregate numbers, the per-scenario evidence for the 26.8% improvement over MT-FTC is missing. Please replace Figure 11 with the actual Solo8 per-scenario statistics or remove the duplicate figure.
minor comments (4)
  1. [Table 6 caption] The caption lists five scenario labels ('Normal', 'Undamaged', 'Sensor-only Damage', 'Detectable Joint Damage', 'Undetectable Joint Damage') for four ratio entries; the labels should match the four Stage-II subcategories defined in Section 3.3.
  2. [Section 4.4, Table 7] The term 'masking value' is ambiguous: the table changes the value placed into zeroed observations, not the additive attention mask value in Eq. (5). Rename the row or clarify the distinction.
  3. [Conclusion] 'restricted motion, weakened motor, or limited velocity' should be 'restricted range of motion, reduced motor force, and limited linear velocity' to match the terminology used throughout the paper.
  4. [Appendix D text] The text says 'Figure 11 is for the Solo8 task (SOTA comparison),' but the figure panels are identical to Figure 9 for G1; this needs correction beyond a caption change.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the reported gains are held-out empirical comparisons, though the paper never isolates the mask contribution via a no-mask ablation.

full rationale

The claimed improvements (36% for transformer, 39% for MLP) are measured on inference settings that differ from training: different terrains, different seeds, different damage timing, and partly different damage ranges (Tables 10-13). The two-stage training and masking mechanism are not defined in terms of the evaluation metric, and no fitted parameter is presented as a prediction. The masking matrix M in Eq. (5) and the zeroed observation V in Eq. (1) are architectural inputs, not re-statements of the success metric. The paper's own ablations (Figure 4, Tables 6-8) test training-stage count, Stage-II sampling ratios, masking values, and curriculum versus stage-based scheduling; they do not remove the attention mask while keeping the two-stage damage distribution. That missing ablation is a confound in the paper's attribution of the gains to the mask encoder, but it is an experimental validity issue, not a circular derivation: the central numbers remain direct comparisons against external baselines (Trf-NM, MLP-NM, BodyTrf, and MT-FTC from Hou et al. 2024). The only self-citations (Qi et al. 2023 and Wan et al. 2024) appear in related-work enumerations and are not load-bearing. Accordingly, no step reduces by construction to its own input, and the circularity score is 2 rather than 0 solely to reflect these minor, non-load-bearing self-citations and the unisolated attribution claim.

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

The central empirical claims rest on environment and training choices rather than on new physics. The Stage II sampling ratio and masked observation value were selected by ablation on the A1 task and then applied to all tasks; damage parameters (ROM, motor, velocity limits) are hand-set per robot. The design also assumes an oracle damage detector for all detectable scenarios, simulation fidelity of IsaacGym, and that zeroing observations plus attention masking removes a damaged joint's influence. No new physical entities are introduced.

free parameters (5)
  • Stage II subcategory sampling ratio = 1:1:1:1
    Selected as the best of seven ratios in an A1-task ablation (Table 6), then fixed as the default across H1, G1, and Solo8; the 5-unit success spread across ratios is 66% to 74%.
  • Masked observation substitution value = 0
    Chosen by ablation on the A1 task (Table 7) because out-of-range values 100 and -100 carried information; the attention mask value itself (minus infinity, Section 3.2.2) was not ablated.
  • Training damage range (number of damaged joints) = 2-4 for A1/H1/G1; 1-3 for Solo8
    Hand-set per task (Tables 10 to 13) to cover the Stage II damage distribution; inference uses a different range (4-5 for A1), so the training choice defines the difficulty of the learned behavior.
  • Joint damage severity limits = ROM 30% of range; motor 4-13 depending on joint group; velocity 3-5
    Environment parameters chosen without sensitivity analysis; G1 has different motor limits per joint (8 hip, 13 knee, 4 ankle, Table 12), and inference uses different severities, so these choices shape the policy's adaptation.
  • PPO loss weights lambda_1 and lambda_2 = Not reported
    Defined in Equation (6) but never given numerically; presumably inherited from Legged Gym defaults (Rudin et al. 2022), which is not stated.
assumptions (5)
  • domain assumption The damage detection module Gamma(O) identifies failed joints correctly in every detectable scenario.
    Section 3.2.1 defines Gamma as producing mask V and attention mask M with no error model, and Appendix B assumes that a detected joint's sensor also fails; a mis-detection would corrupt the observation and attention inputs.
  • domain assumption IsaacGym physics faithfully captures damaged quadruped and humanoid locomotion.
    All quantitative claims come from IsaacGym simulation (Section 4.1); the introduction and conclusion claim real-world applicability, but no hardware transfer is performed, so sim-to-real fidelity is assumed.
  • domain assumption Zeroing a damaged joint's observation and masking its attention removes its influence on the policy.
    Sections 3.2.1 and 3.2.2 assume masked inputs and minus-infinity attention entries make the network stop relying on malfunctioning limbs; for undetectable damage no mask is applied, and adaptation must come from dynamic state alone.
  • domain assumption Distance reached without falling is a valid proxy for task completion.
    Section 4.2 defines success as crossing 1-5 unit radii from the damaged position; this self-defined metric is assumed cumulative, an assumption that Table 2 (MLP-NM row) contradicts.
  • standard math PPO with the Legged Gym default setup is an adequate training substrate for both baselines and UMC.
    The paper relies on PPO (Schulman et al. 2017) and Legged Gym (Rudin et al. 2022) without modification or convergence analysis; the one-stage comparison in Figure 4 stops at 2500 iterations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UMC: Unified Resilient Controller for Legged Robots with Joint Malfunctions." pith.science (2026). https://pith.science/paper/6WDPNJ7V

@misc{pith2026250203035,
  author       = {Pith},
  title        = {Pith review of: UMC: Unified Resilient Controller for Legged Robots with Joint Malfunctions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6WDPNJ7V}},
  note         = {Machine review of arXiv:2502.03035}
}
read the original abstract

Adaptation to unpredictable damages is crucial for autonomous legged robots, yet existing methods based on multi-policy or meta-learning frameworks face challenges like limited generalization and complex maintenance. To address this issue, we first analyze and summarize eight types of damage scenarios, including sensor failures and joint malfunctions. Then, we propose a novel, model-free, two-stage training framework, Unified Malfunction Controller (UMC), incorporating a masking mechanism to enhance damage resilience. Specifically, the model is initially trained with normal environments to ensure robust performance under standard conditions. In the second stage, we use masks to prevent the legged robot from relying on malfunctioning limbs, enabling adaptive gait and movement adjustments upon malfunction. Experimental results demonstrate that our approach improves the task completion capability by an average of 36% for the transformer and 39% for the MLP across three locomotion tasks. The source code and trained models will be made available to the public.

Figures

Figures reproduced from arXiv: 2502.03035 by the authors.

Figure 1
Figure 1. Qualitative and Quantitative Comparison of Our UMC Framework with Baselines and a SOTA Method. ‘Trf’ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. UMC system for transformer-based Actor-Model Architecture. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative Comparison Between Methods Under Damaged Scenarios. ‘Baseline’ refers to robots trained using [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Comparison of One-Stage and Two-Stage Training [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Demonstration of different damage conditions. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: UMC Framework for MLP-based Actor-Model Architecture. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Performance of Five Methods Under Different Damage Conditions in the A1-Walk Task. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Performance of Five Methods Under Different Damage Conditions in the Unitree-H1 Task. [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Performance of Five Methods Under Different Damage Conditions in the Unitree-G1 Task. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Average Performance of Five Methods Under Different Damage Conditions Across Three Tasks. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: Performance Between UMC and ‘MT-FTC’ Under Different Damage Conditions in the Solo8 Task. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: Full Comparison for the Ratio and Masking-Value Ablations Under Different Damage Conditions. [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 17 canonical work pages

  1. [1]

    and Jisha, V

    A, A. and Jisha, V. R. Reinforcement learning based control of a quadruped robot. 2022 IEEE 19th India Council International Conference (INDICON), pp.\ 1--6, 2022. URL https://api.semanticscholar.org/CorpusID:256945713

    work page 2022

  2. [2]

    Advances in real‐world applications for legged robots

    Bellicoso, D., Bjelonic, M., Wellhausen, L., Holtmann, K., G \"u nther, F., Tranzatto, M., Fankhauser, P., and Hutter, M. Advances in real‐world applications for legged robots. Journal of Field Robotics, 35: 0 1311 -- 1326, 2018. URL https://api.semanticscholar.org/CorpusID:117305362

    work page 2018

  3. [3]

    Fast adaptation dynamics model for robot’s damage recovery

    Chen, C., Wang, D., Yu, J., Xiang, P., Lu, H., Wang, Y., and Xiong, R. Fast adaptation dynamics model for robot’s damage recovery. In 2022 IEEE International Conference on Real-time Computing and Robotics (RCAR), pp.\ 45--50, 2022. doi:10.1109/RCAR54675.2022.9872230

    work page arXiv 2022

  4. [4]

    Meta reinforcement learning of locomotion policy for quadruped robots with motor stuck

    Chen, C., Li, C., Lu, H., Wang, Y., and Xiong, R. Meta reinforcement learning of locomotion policy for quadruped robots with motor stuck. IEEE Transactions on Automation Science and Engineering, pp.\ 1--15, 2024. doi:10.1109/TASE.2024.3424328

    work page arXiv 2024

  5. [5]

    Decision transformer: Reinforcement learning via sequence modeling, 2021

    Chen, L., Lu, K., Rajeswaran, A., Lee, K., Grover, A., Laskin, M., Abbeel, P., Srinivas, A., and Mordatch, I. Decision transformer: Reinforcement learning via sequence modeling, 2021. URL https://arxiv.org/abs/2106.01345

    arXiv 2021

  6. [6]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. URL https://arxiv.org/abs/2010.11929

    arXiv 2021

  7. [7]

    An open torque-controlled modular robot architecture for legged locomotion research

    Grimminger, F., Meduri, A., Khadiv, M., Viereck, J., Wüthrich, M., Naveau, M., Berenz, V., Heim, S., Widmaier, F., Flayols, T., Fiene, J., Badri-Spröwitz, A., and Righetti, L. An open torque-controlled modular robot architecture for legged locomotion research. IEEE Robotics and Automation Letters, 5 0 (2): 0 3650--3657, 2020. doi:10.1109/LRA.2020.2976639

    work page arXiv 2020

  8. [8]

    Fault self-diagnosis for modular robotic systems using m-lattice modules

    Guan, E., Fei, J., Pan, G., Fu, Z., Yan, W., and Zhao, Y. Fault self-diagnosis for modular robotic systems using m-lattice modules. International Journal of Advanced Robotic Systems, 12, 2015. URL https://api.semanticscholar.org/CorpusID:43325716

    work page 2015

Show all 34 references
  1. [9]

    Decentralized motor skill learning for complex robotic systems, 2023

    Guo, Y., Jiang, Z., Wang, Y.-J., Gao, J., and Chen, J. Decentralized motor skill learning for complex robotic systems, 2023. URL https://arxiv.org/abs/2306.17411

    2023 arXiv

  2. [10]

    Metamorph: Learning universal controllers with transformers, 2022

    Gupta, A., Fan, L., Ganguli, S., and Fei-Fei, L. Metamorph: Learning universal controllers with transformers, 2022. URL https://arxiv.org/abs/2203.11931

    2022 arXiv

  3. [11]

    Structure-aware transformer policy for inhomogeneous multi-task reinforcement learning

    Hong, S., Yoon, D., and Kim, K.-E. Structure-aware transformer policy for inhomogeneous multi-task reinforcement learning. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=fy_XRVHqly

    2022

  4. [12]

    Multi-task learning of active fault-tolerant controller for leg failures in quadruped robots, 2024

    Hou, T., Tu, J., Gao, X., Dong, Z., Zhai, P., and Zhang, L. Multi-task learning of active fault-tolerant controller for leg failures in quadruped robots, 2024. URL https://arxiv.org/abs/2402.08996

    2024 arXiv

  5. [13]

    u nther, F., Bellicoso, D., Tsounis, V., Fankhauser, P., Diethelm, R., Bachmann, S., Bl \

    Hutter, M., Gehring, C., Lauber, A., G \"u nther, F., Bellicoso, D., Tsounis, V., Fankhauser, P., Diethelm, R., Bachmann, S., Bl \"o sch, M., Kolvenbach, H., Bjelonic, M., Isler, L., and Meyer, K. Anymal - toward legged robots for harsh environments. Advanced Robotics, 31: 0 9...

    2017

  6. [14]

    A study of self-diagnosis system of an autonomous mobile robot: expansion of state sensory systems

    Kawabata, K., Akamatsu, T., and Asama, H. A study of self-diagnosis system of an autonomous mobile robot: expansion of state sensory systems. In IEEE/RSJ International Conference on Intelligent Robots and Systems, volume 2, pp.\ 1802--1807 vol.2, 2002. doi:10.1109/IRDS.2002.1044017

    2002 arXiv

  7. [15]

    Map-based multi-policy reinforcement learning: Enhancing adaptability of robots by deep reinforcement learning, 2017

    Kume, A., Matsumoto, E., Takahashi, K., Ko, W., and Tan, J. Map-based multi-policy reinforcement learning: Enhancing adaptability of robots by deep reinforcement learning, 2017. URL https://arxiv.org/abs/1710.06117

    2017 arXiv

  8. [16]

    My body is a cage: the role of morphology in graph-based incompatible control, 2021

    Kurin, V., Igl, M., Rocktäschel, T., Boehmer, W., and Whiteson, S. My body is a cage: the role of morphology in graph-based incompatible control, 2021. URL https://arxiv.org/abs/2010.01856

    2021 arXiv

  9. [17]

    Isaac gym: High performance gpu-based physics simulation for robot learning, 2021

    Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., and State, G. Isaac gym: High performance gpu-based physics simulation for robot learning, 2021. URL https://arxiv.org/abs/2108.10470

    2021 arXiv

  10. [18]

    S., Abbeel, P., Levine, S., and Finn, C

    Nagabandi, A., Clavera, I., Liu, S., Fearing, R. S., Abbeel, P., Levine, S., and Finn, C. Learning to adapt in dynamic, real-world environments through meta-reinforcement learning, 2019. URL https://arxiv.org/abs/1803.11347

    2019 arXiv

  11. [19]

    and Samanta, B

    Nwaonumah, E. and Samanta, B. Deep reinforcement learning for visual navigation of wheeled mobile robots. In 2020 SoutheastCon, pp.\ 1--8, 2020. doi:10.1109/SoutheastCon44009.2020.9249654

    2020

  12. [20]

    High-quality entity segmentation

    Qi, L., Kuen, J., Guo, W., Shen, T., Gu, J., Jia, J., Lin, Z., and Yang, M.-H. High-quality entity segmentation. In ICCV, 2023

    2023

  13. [21]

    Quamar, M. M. and Nasir, A. Review on fault diagnosis and fault-tolerant control scheme for robotic manipulators: Recent advances in ai, machine learning, and digital twin, 2024. URL https://arxiv.org/abs/2402.02980

    2024 arXiv

  14. [22]

    Humanoid locomotion as next token prediction, 2024

    Radosavovic, I., Zhang, B., Shi, B., Rajasegaran, J., Kamat, S., Darrell, T., Sreenath, K., and Malik, J. Humanoid locomotion as next token prediction, 2024. URL https://arxiv.org/abs/2402.19469

    2024 arXiv

  15. [23]

    Fast adaptation via policy-dynamics value functions, 2020

    Raileanu, R., Goldstein, M., Szlam, A., and Fergus, R. Fast adaptation via policy-dynamics value functions, 2020. URL https://arxiv.org/abs/2007.02879

    2020 arXiv

  16. [24]

    Learning to walk in minutes using massively parallel deep reinforcement learning, 2022

    Rudin, N., Hoeller, D., Reist, P., and Hutter, M. Learning to walk in minutes using massively parallel deep reinforcement learning, 2022. URL https://arxiv.org/abs/2109.11978

    2022 arXiv

  17. [25]

    Proximal policy optimization algorithms, 2017

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347

    2017 arXiv

  18. [26]

    Body transformer: Leveraging robot embodiment for policy learning, 2024

    Sferrazza, C., Huang, D.-M., Liu, F., Lee, J., and Abbeel, P. Body transformer: Leveraging robot embodiment for policy learning, 2024. URL https://arxiv.org/abs/2408.06316

    2024 arXiv

  19. [27]

    Learning to combine primitive skills: A step towards versatile robotic manipulation, 2020

    Strudel, R., Pashevich, A., Kalevatykh, I., Laptev, I., Sivic, J., and Schmid, C. Learning to combine primitive skills: A step towards versatile robotic manipulation, 2020. URL https://arxiv.org/abs/1908.00722

    2020 arXiv

  20. [28]

    Learning agile locomotion via adversarial training, 2020

    Tang, Y., Tan, J., and Harada, T. Learning agile locomotion via adversarial training, 2020. URL https://arxiv.org/abs/2008.00603

    2020 arXiv

  21. [29]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762

    2023 arXiv

  22. [30]

    Vint-6d: A large-scale object-in-hand dataset from vision, touch and proprioception

    Wan, Z., Ling, Y., Yi, S., Qi, L., Lee, W., Lu, M., Yang, S., Teng, X., Lu, P., Yang, X., et al. Vint-6d: A large-scale object-in-hand dataset from vision, touch and proprioception. In ICML, 2024

    2024

  23. [31]

    M., Posa, M., Hu, Y., Escande, A., Mansard, N., and Prete, A

    Wensing, P. M., Posa, M., Hu, Y., Escande, A., Mansard, N., and Prete, A. D. Optimization-based control for dynamic legged robots, 2022. URL https://arxiv.org/abs/2211.11644

    2022 arXiv

  24. [32]

    Fault-aware robust control via adversarial reinforcement learning

    Yang, F., Yang, C., Guo, D., Liu, H., and Sun, F. Fault-aware robust control via adversarial reinforcement learning. In 2021 IEEE 11th Annual International Conference on CYBER Technology in Automation, Control, and Intelligent Systems (CYBER), pp.\ 109--115, 2021. doi:10.1109/...

    2021

  25. [33]

    Robot parkour learning

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

    2023

  26. [34]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.