Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Rapid Embodiment Adaptation for Quadrupedal Locomotion

T0 review · 3 major / 5 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 A quadruped re-identifies its changed joint limits and trunk mass within 0.4 seconds and keeps walking under a locked leg and a 5 kg payload.

desk verdict A solid incremental RMA-style adaptation paper with a clean sim pipeline and promising real trials, but the hardware identification claim is not directly measured and the 'fully locked leg' overstates a 0.1-scale virtual clamp. read the letter →

arxiv 2608.01506 v1 pith:F2IZ2JHC submitted 2026-08-02 cs.RO cs.AIcs.LG

Dichen Li , Bo Ai , Nico Bohlinger , Jan Peters , Hao Su , Henrik I. Christensen This is my paper
classification cs.ROcs.AIcs.LG
keywords embodimentadaptationquadrupedallocomotiononlinesystemidentificationjoint-limitconstraintspayloadmassvariationcross-embodimentpolicysim-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 sets out to solve a practical gap: learning-based locomotion policies are trained for a known, fixed body, so they fail when the robot's hardware changes—a joint starts to lock, or a payload is strapped on. Its proposal is to treat the robot's own body as a hidden variable: a lightweight adaptation module reads the last 20 timesteps of sensor-action history, estimates physical parameters such as per-joint range limits and trunk mass, and hands those estimates to a generalist policy that was trained across randomized versions of these parameters. The result, if correct, is that a quadruped can re-identify its new body within about 0.4 seconds of interaction and adjust its gait accordingly. In simulation the estimated parameters match ground truth and closed-loop control outperforms an end-to-end policy that maps histories directly to actions; on a real Unitree Go2 the approach keeps the robot walking with a heavily restricted front leg and a 5 kg payload, where a non-adaptive baseline falls. The authors position this as evidence that explicit online embodiment identification is practical and a step toward controllers that handle changing or degraded hardware.

What carries the argument

The load-bearing mechanism is the two-stage training split: first, a cross-embodiment base policy is trained by reinforcement learning with ground-truth embodiment descriptions and aggressive randomization of joint-limit scales and trunk masses, so it learns to move well under any embodiment it is told about; then an adaptation module—an LSTM temporal encoder feeding a joint-wise Transformer—is trained by supervised regression to recover the same embodiment parameters (per-joint nominal position, torque limit, range endpoints, plus global trunk mass, or a learned latent embedding) from a sliding window of 20 observation-action pairs. At deployment the base policy is frozen and conditioned on

What would settle it

Physically lock one leg joint with a rigid clamp instead of the high-gain virtual constraint, run the same pipeline, and check whether the adaptation module's predicted joint-limit scale converges to the simulated equivalent within 0.4 seconds and whether the conditioned policy keeps the robot upright. A consistent bias when comparing predicted trunk-mass offset against a series of known weights (1, 2, 3, 4, 5 kg) on hardware would likewise show the identification is not physically grounded.

Watch

Extended reading notes

Core claim

The central claim is that decomposing adaptation into (1) supervised inference of embodiment parameters from a short interaction window and (2) conditioning a frozen embodiment-conditioned policy on that inference yields rapid, robust adaptation to two representative embodiment shifts: joint-limit scaling and trunk-mass offset. The insight is that the policy never needs to infer the body from scratch by end-to-end learning; it already knows how to move with any embodiment it can be told about, so the only missing piece is a fast, explicit estimate of which embodiment is currently present. With a 20-timestep window at 50 Hz, the adaptation module converges to the true parameters within 0.4 se

Load-bearing premise

The load-bearing premise is that a virtual joint restriction (a high-gain controller refusing to move outside a fixed range) and a bolted-on payload produce the same interaction signatures as the joint-limit and mass parameters sampled during simulation; if that mapping is wrong, the inferred estimates have no meaning on the real robot and the frozen policy would not transfer.

Editorial extensions

If this is right

  • If the central claim holds, a robot controller no longer needs a known, static body model at deployment: it can re-estimate joint-range and mass changes from a few steps of interaction and keep walking.
  • Because adaptation converges in roughly 0.4 seconds, abrupt mid-episode failures—a leg locking or a payload dropped onto the trunk—can be compensated online rather than requiring a retraining or calibration phase.
  • The explicit parameter estimates are interpretable: the robot visibly switches to an asymmetric 'crutch' gait when a leg is restricted and to a lower, longer-stance gait under load, suggesting the adaptation is grounded in physical causes rather than opaque history matching.
  • Outperforming the implicit end-to-end baseline in simulation indicates that explicitly predicting embodiment parameters can be faster and more sample-efficient at deployment than long-context history-to-action policies.

Reading between the lines

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

  • Because the real-world joint restriction is implemented by a high-gain virtual constraint rather than a physical lock, the strongest version of the claim would need a test with a mechanically jammed joint; a physical jam adds friction, backlash, and possible collisions that may not match the simulated joint-limit signature.
  • The same two-stage supervised identification recipe could be applied to other embodiment parameters not covered here—actuator torque limits, foot-ground friction, or inertia tensors—provided the simulator randomizes them and the base policy is trained on them.
  • The explicit estimates double as a diagnostic signal: the estimated mass offset and joint-limit scale over time could flag mechanical wear or payload shifts before the robot falls, a maintenance use case the paper does not develop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage framework for online embodiment adaptation in quadrupedal locomotion. A cross-embodiment base policy (URMA-based) is trained with embodiment randomization and oracle conditioning, while a separate adaptation module is trained via supervised regression to infer either explicit physical embodiment parameters (per-joint limit scale, torque limit, range, nominal position, and trunk mass) or a latent embodiment embedding from a short (≈20 timestep) history of proprioceptive observations and past actions. At deployment, the frozen base policy conditions on the estimated parameters. Experiments in IsaacLab compare architectures, history length, and closed-loop sweeps against no-adaptation, oracle, and implicit end-to-end baselines. Real-world experiments on a Unitree Go2 evaluate static and dynamic joint-limit scaling (implemented by a high-gain virtual clamp) and payload addition, reporting normalized success, time-to-failure, distance, and heading-deviation metrics over 10 trials. The paper claims 0.4-second identification, near-oracle simulation performance, and stable real-world locomotion under a 'fully locked leg' and a 5 kg payload.

Significance. If fully supported, the framework would be a useful step toward practical online identification of robot-side kinematic and dynamic changes: the decomposition into a supervised estimator and a frozen embodiment-conditioned policy is clean, and the explicit/latent comparison is informative. The simulation evidence is coherent and reasonably thorough: the adaptation converges quickly, closed-loop performance approaches the oracle, and the comparison against an implicit adaptation baseline is a relevant control. The real-robot trials, while limited in number and metric detail, show consistent improvements over no-adaptation. The key weakness is that the paper's strongest advertised claims—hardware identification and a 'fully locked leg'—are not directly established by the reported hardware data, and the simulator-to-hardware mapping of the joint constraint is unverified. These are load-bearing for the central contribution, but they are addressable with additional measurements or calibrated language.

major comments (3)
  1. [§V-C, Fig. 6, Table IV] The central hardware claim—that the adaptation module identifies embodiment parameters online—is not supported by any direct measurement. Table IV reports only locomotion outcomes (success, TTF, distance, heading deviation), and Fig. 6 shows gait snapshots. The statement in §V-C that 'Predictions converge within approximately 20 timesteps' is not accompanied by hardware plots of the predicted joint-limit scale or mass. Without logged φ̂_t, the observed improvement over No Adaptation cannot be attributed to correct embodiment identification; it could come from the base policy's general robustness or from a correlated non-causal cue. Please either (a) provide hardware plots of estimated embodiment parameters (and ground truth where available) for the dynamic trials, or (b) restrict the real-world conclusion to closed-loop outcome improvements and keep identification claims in simulation.
  2. [Abstract; §V-C, Table IV] The abstract and §VI describe the hardest hardware condition as a 'fully locked leg,' but Table IV's hardest joint condition is 'FR Joint Lock 0.1,' meaning the joint range is scaled to 10% of nominal, and the constraint is implemented by a high PD gain outside the prescribed range (§IV-A), not by a physical lock. A 10%-range virtual clamp is severe, but it is not a fully locked leg. Moreover, in that condition Ours has normalized success 0.375, which the table's legend maps to roughly 30% success across 10 trials, so 'maintains stable locomotion' overstates reliability. Please characterize the condition accurately, report absolute success counts/rates, and, if 'fully locked' is to be claimed, add a zero-range or mechanical-lock condition with corresponding metrics.
  3. [§IV-A; §V-C] The simulator-to-hardware transfer of the joint constraint is not validated. The policy and adaptation module are trained on randomized 'joint limit scale' in IsaacLab, whereas hardware constraints are imposed by 'applying a high control gain outside a prescribed joint range' (§IV-A). These two implementations can produce different interaction signatures (e.g., hard kinematic stops versus torque/PD saturation), so the estimated φ̂_t on hardware could lie outside the training distribution even if the resulting behavior looks similar. Because φ̂_t is never measured on hardware, this potential mismatch is untested. Please either validate the mapping—e.g., by comparing simulator and real-world state-action histories or predicted φ̂_t under matched constraints—or add a simulation ablation with a virtual high-gain clamp to show the estimator is robust to constraint implementation details.
minor comments (5)
  1. [Eq. (1)] The notation '⟨ρ_t, m_t⟩' is ambiguous; it appears to denote concatenation of the embodiment descriptor and mass, but the angle brackets are unusual. Please define the target vector explicitly.
  2. [Table IV] The normalization description is confusing: 'scaled relative to the best observed value' followed by 'A score of 1.0 corresponds to a success rate of 80%, TTF of 13 s, distance of 2.4 m, and heading deviation of 3°' needs clarification. Report raw absolute values or a clear per-column normalization key.
  3. [Fig. 4] The choice of a 20-step window is justified as a trade-off, but Fig. 4 shows that 50 steps yield lower RMSE. Please provide closed-loop results for the 50-step window or explain why 20 is preferable beyond prediction RMSE.
  4. [§V-B] The 'Implicit End-to-End Policy' baseline is described only briefly. Please specify its input history length, training data, and whether it received the same embodiment randomization, so the comparison is reproducible.
  5. [Throughout] Minor typos include 'paramters' in §III-C and the ungrammatical sentence in §IV-A ending with '..' Please proofread.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: adaptation module is supervised against simulator ground truth and evaluated on external locomotion metrics.

full rationale

The derivation chain is self-contained. The base policy π_base is trained by RL with oracle embodiment information, and the adaptation module f is trained separately by supervised regression (Eq. 1) against recorded simulator ground-truth adaptable parameters (d_t, m_t). Deployment conditions a frozen policy on f(h_t). The training target is the simulator's actual embodiment parameter, not the downstream return, so the parameter estimates are not fitted to the evaluation metric. Closed-loop evaluation uses external locomotion outcomes (return, episode length, success rate, TTF, distance, heading deviation) that do not appear in the adaptation module's loss. The S_rel normalization (Eq. 2) is explicitly by construction a 0–100% rescaling between No Adaptation and Oracle; it is a reporting convention, not a derived prediction or a claim that the method closes the gap by definition. The cited prior work [3,4,6] overlaps with the authors and provides the URMA architecture and training scheme, but the central claim—that a short interaction history suffices to identify joint-limit scale and trunk mass offset—is supported by direct parameter-error measurements (Table III, Fig. 5) against simulator ground truth and by comparisons to No Adaptation and Oracle baselines. No uniqueness theorem or ansatz is imported via self-citation in a load-bearing way. The hardware evaluation does not directly measure φ̂_t and implements joint locks as a virtual high-gain clamp; this is a sim-to-real evidence gap, not a circularity, and does not affect the circularity score.

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

No new physical entities are posited. The latent embodiment representation is a learned feature vector, not a physical object; the explicit representation consists of ordinary robot parameters. The central free parameters are training and architecture choices, not derived quantities.

free parameters (5)
  • History window length T = 20 timesteps (0.4 s at 50 Hz)
    Hand-selected from Fig. 4 trade-off curve; the headline '0.4 s adaptation' claim depends on this choice.
  • Joint-limit scale randomization range = [0, 0.5]
    Training-time range chosen by hand; the policy is tested at nominal scale 1.0, i.e., outside the sampled training range.
  • Trunk mass offset randomization range = [-3.0 kg, +7.0 kg]
    Hand-chosen range; real 5.0 kg payload falls inside it.
  • Embodiment configuration grid = 36 discrete configurations
    Offline dataset samples only 36 configurations; continuous generalization claims depend on interpolation over this grid.
  • Latent representation dimension d_z = Not reported in text
    Hyperparameter of the latent variant, never specified, so the latent experiment is not fully reproducible.
assumptions (3)
  • domain assumption IsaacLab simulation dynamics faithfully represent the Unitree Go2, including the mapping from joint-limit scale and trunk-mass offset to real-world behavior.
    The base policy and adaptation module are trained entirely in simulation; the real-world section (V-C) assumes these parameters transfer without fine-tuning.
  • domain assumption The embodiment parameter vector (per-joint ranges, torque limits, nominal positions, trunk mass) is a sufficient control-conditioning variable.
    The method conditions a frozen policy on estimated phi_t; if the chosen parameterization omits dynamics that matter for control, the conditioning cannot recover performance (Section III-B, III-C).
  • ad hoc to paper A 20-step sliding window of proprioceptive observations and prior actions suffices to identify the embodiment parameters at control-relevant accuracy.
    Fig. 4 shows accuracy improves until 50 steps; the paper still chooses 20 as a practical trade-off, making the 0.4 s convergence claim dependent on this window.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rapid Embodiment Adaptation for Quadrupedal Locomotion." pith.science (2026). https://pith.science/paper/F2IZ2JHC

@misc{pith2026260801506,
  author       = {Pith},
  title        = {Pith review of: Rapid Embodiment Adaptation for Quadrupedal Locomotion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2IZ2JHC}},
  note         = {Machine review of arXiv:2608.01506}
}
read the original abstract

Humans readily adapt their movements as their bodies change through aging, injury, or load carrying, but learning-based robot policies often break when hardware properties shift. We introduce an online embodiment adaptation framework for quadrupedal locomotion that infers embodiment parameters from short interaction histories and conditions control on the inferred hardware state. Our method pairs a generalist policy trained under embodiment randomization with a lightweight adaptation module that identifies physical changes within half a second. We evaluate two representative forms of embodiment variation: joint-range constraints and trunk-mass changes, corresponding to joint-level kinematic degradation and body-level dynamic variation. In simulation, the module accurately estimates these changes and enables closed-loop control that substantially outperforms policies conditioned directly on interaction history. On a real Unitree Go2 robot, our system maintains stable locomotion under severe instances of the evaluated changes, including a fully locked leg and a 5 kg payload, where non-adaptive methods fail. These results demonstrate the practicality of explicit online embodiment identification for rapid adaptation to joint-limit and payload-mass changes, and provide a step toward handling broader forms of uncertain, degraded, or changing robot hardware.

Figures

Figures reproduced from arXiv: 2608.01506 by the authors.

Figure 1
Figure 1. Pipeline of the proposed online embodiment adaptation framework. The cross-embodiment base policy πbase is trained to handle locomotion tasks across different embodiments with the ground-truth embodiment description. The adaptation module f is trained to infer embodiment parameters from observation history for the base policy to condition on. The above adaptation method (Explicit Representation) infers physical embo… view at source ↗
Figure 2
Figure 2. Mean episode return and episode length versus joint limit scaling and trunk mass offset in simulation. Raw values; max 1000 steps = 20 s. We evaluate under (a) varying joint limit scales and (b) varying trunk mass offset. 0.0 0.2 0.4 0.6 0.8 1.0 Joint Limit Scale 50 0 50 100 150 Relative Perf (%) 5 0 5 10 15 Mass Offset (kg) 50 0 50 100 150 Relative Perf (%) (a) Joint Limit Scaling (b) Trunk Mass Offset No Adapt (0%… view at source ↗
Figure 5
Figure 5. Qualitative results on embodiment parameter prediction under 2 abrupt changes: (a) joint limit change; (b) payload mass change. Our adaptation module is able to identify changing parameters within 0.4 seconds from interaction history. B. Closed-Loop Evaluation in Simulation We evaluate the robustness of locomotion under controlled embodiment changes. Joint limits are scaled relative to the original per-joint limits.… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Identification RMSE versus history window length on the validation set. achieves the best performance and is used for the rest of the paper. History length [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Qualitative results for online adaptation under embodiment changes during deployment. Our adaptation module identifies embodiment parameters within half a second, and the policy conditioned on the inferred parameters adjusts the gait to account for the changed embodime…
Figure 7
Figure 7. Figure 7: Gait comparison between our online adaptation (Explicit Representation) method and the No Adaptation baseline under embodiment changes. We evaluate two embodiment changes: (a) restricting the front-right leg joint limits to 0.3 of their nominal range and (b) adding a 5…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

  1. [1]

    A Review of Learning-Based Dynamics Models for Robotic Manipulation

    Bo Ai et al. “A Review of Learning-Based Dynamics Models for Robotic Manipulation”. In:Science Robotics(2025).DOI: 10.1126/scirobotics.adt1497

    doi:10.1126/scirobotics.adt1497 2025

  2. [2]

    RoboPack: Learning Tactile-Informed Dynam- ics Models for Dense Packing

    Bo Ai et al. “RoboPack: Learning Tactile-Informed Dynam- ics Models for Dense Packing”. In:Robotics: Science and Systems (RSS)(2024)

    work page 2024

  3. [3]

    Towards Embodiment Scaling Laws in Robot Locomotion

    Bo Ai et al. “Towards Embodiment Scaling Laws in Robot Locomotion”. In:Conference on Robot Learning (CoRL) (2025)

    work page 2025

  4. [4]

    Multi-Embodiment Loco- motion at Scale with extreme Embodiment Randomization

    Nico Bohlinger and Jan Peters. “Multi-Embodiment Loco- motion at Scale with extreme Embodiment Randomization”. In:arXiv preprint arXiv:2509.02815(2025)

    arXiv 2025

  5. [5]

    Gait in Eight: Efficient On-Robot Learning for Omnidirectional Quadruped Locomotion

    Nico Bohlinger et al. “Gait in Eight: Efficient On-Robot Learning for Omnidirectional Quadruped Locomotion”. In: International Conference on Intelligent Robots and Systems (2025)

    work page 2025

  6. [6]

    One Policy to Run Them All: an End-to-end Learning Approach to Multi-Embodiment Loco- motion

    Nico Bohlinger et al. “One Policy to Run Them All: an End-to-end Learning Approach to Multi-Embodiment Loco- motion”. In:Conference on Robot Learning(2024)

    work page 2024

  7. [7]

    Expressive whole-body control for hu- manoid robots

    Xuxin Cheng et al. “Expressive whole-body control for hu- manoid robots”. In:arXiv preprint arXiv:2402.16796(2024)

    arXiv 2024

  8. [8]

    Extreme Parkour with Legged Robots

    Xuxin Cheng et al. “Extreme Parkour with Legged Robots”. In:RoboLetics: Workshop on robot learning in athletics @ CoRL. 2023

    work page 2023

Show all 38 references
  1. [9]

    Learning quadrupedal locomotion on deformable terrain

    Suyoung Choi et al. “Learning quadrupedal locomotion on deformable terrain”. In:Science Robotics8.74 (2023), eade2256

    2023

  2. [10]

    Empirical evaluation of gated re- current neural networks on sequence modeling

    Junyoung Chung et al. “Empirical evaluation of gated re- current neural networks on sequence modeling”. In:arXiv preprint arXiv:1412.3555(2014)

    2014 arXiv

  3. [11]

    Metamorph: learning universal con- trollers with transformers

    Agrim Gupta et al. “Metamorph: learning universal con- trollers with transformers”. In:International Conference on Learning Representations. ICLR. 2022

    2022

  4. [12]

    Scaling Cross-Embodiment World Mod- els for Dexterous Manipulation

    Zihao He et al. “Scaling Cross-Embodiment World Mod- els for Dexterous Manipulation”. In:arXiv preprint arXiv:2511.01177(2025)

    2025 arXiv

  5. [13]

    One policy to control them all: Shared modular policies for agent- agnostic control

    Wenlong Huang, Igor Mordatch, and Deepak Pathak. “One policy to control them all: Shared modular policies for agent- agnostic control”. In:International Conference on Machine Learning. PMLR. 2020, pp. 4455–4464

    2020

  6. [14]

    Adapting rapid motor adaptation for bipedal robots

    Ashish Kumar et al. “Adapting rapid motor adaptation for bipedal robots”. In:2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE. 2022, pp. 1161–1168

    2022

  7. [15]

    Rma: Rapid motor adaptation for legged robots

    Ashish Kumar et al. “Rma: Rapid motor adaptation for legged robots”. In: 2021

    2021

  8. [16]

    Loco- Former: Generalist Locomotion via Long-context Adapta- tion

    Min Liu, Deepak Pathak, and Ananye Agarwal. “Loco- Former: Generalist Locomotion via Long-context Adapta- tion”. In:Conference on Robot Learning. PMLR. 2025, pp. 532–546

    2025

  9. [17]

    Tools for the body (schema)

    Angelo Maravita and Atsushi Iriki. “Tools for the body (schema)”. In:Trends in Cognitive Sciences8.2 (2004), pp. 79–86.ISSN: 1364-6613.DOI:https://doi.org/ 10 . 1016 / j . tics . 2003 . 12 . 008.URL:https : //www.sciencedirect.com/science/article/ pii/S1364661303003450

    2004

  10. [18]

    Learning robust perceptive locomotion for quadrupedal robots in the wild

    Takahiro Miki et al. “Learning robust perceptive locomotion for quadrupedal robots in the wild”. In:Science robotics7.62 (2022), eabk2822

    2022

  11. [19]

    Isaac lab: A gpu-accelerated simula- tion framework for multi-modal robot learning

    Mayank Mittal et al. “Isaac lab: A gpu-accelerated simula- tion framework for multi-modal robot learning”. In:arXiv preprint arXiv:2511.04831(2025)

    2025 arXiv

  12. [20]

    Long-term adaptation to dynamics of reaching movements: a PET study

    R Nezafat, R Shadmehr, and H H Holcomb. “Long-term adaptation to dynamics of reaching movements: a PET study”. en. In:Experimental Brain Research140.1 (Sept. 2001), pp. 66–76

    2001

  13. [21]

    Real-world embodied AI through a morphologically adaptive quadruped robot

    Tønnes F Nygaard et al. “Real-world embodied AI through a morphologically adaptive quadruped robot”. In:Nature Machine Intelligence3.5 (2021), pp. 410–419

    2021

  14. [22]

    GET-Zero: Graph Embod- iment Transformer for Zero-shot Embodiment Generaliza- tion

    Austin Patel and Shuran Song. “GET-Zero: Graph Embod- iment Transformer for Zero-shot Embodiment Generaliza- tion”. In:2025 IEEE International Conference on Robotics and Automation (ICRA). 2025

    2025

  15. [23]

    Get-zero: Graph embodi- ment transformer for zero-shot embodiment generalization

    Austin Patel and Shuran Song. “Get-zero: Graph embodi- ment transformer for zero-shot embodiment generalization”. In:2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE. 2025, pp. 14262–14269

    2025

  16. [24]

    π 0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities

    Physical Intelligence et al. “π 0.7: A Steerable Generalist Robotic Foundation Model with Emergent Capabilities”. In: arXiv preprint arXiv:2604.15483(2026)

    2026 arXiv

  17. [25]

    In-Hand Object Rotation via Rapid Motor Adaptation

    Haozhi Qi et al. “In-Hand Object Rotation via Rapid Motor Adaptation”. In:Conference on Robot Learning (CoRL). 2022

    2022

  18. [26]

    Learning to walk in minutes using mas- sively parallel deep reinforcement learning

    Nikita Rudin et al. “Learning to walk in minutes using mas- sively parallel deep reinforcement learning”. In:Conference on robot learning. 2022

    2022

  19. [27]

    Long short- term memory

    J ¨urgen Schmidhuber, Sepp Hochreiter, et al. “Long short- term memory”. In:Neural Comput9.8 (1997), pp. 1735– 1780

    1997

  20. [28]

    Proximal policy optimization algo- rithms

    John Schulman et al. “Proximal policy optimization algo- rithms”. In:arXiv preprint arXiv:1707.06347(2017)

    2017 arXiv

  21. [29]

    Body Transformer: Leveraging Robot Embodiment for Policy Learning

    Carmelo Sferrazza et al. “Body Transformer: Leveraging Robot Embodiment for Policy Learning”. In:Conference on Robot Learning. PMLR. 2025, pp. 3407–3424

    2025

  22. [30]

    Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations

    Maximilian Stasica et al. “Bridge the Gap: Enhancing Quadruped Locomotion with Vertical Ground Perturbations”. In:International Conference on Intelligent Robots and Sys- tems(2025)

    2025

  23. [31]

    Accurate power consumption es- timation method makes walking robots energy efficient and quiet

    Giorgio Valsecchi et al. “Accurate power consumption es- timation method makes walking robots energy efficient and quiet”. In:2024 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS). IEEE. 2024, pp. 13282– 13288

    2024

  24. [32]

    Attention is all you need

    Ashish Vaswani et al. “Attention is all you need”. In:Ad- vances in neural information processing systems30 (2017)

    2017

  25. [33]

    Nervenet: Learning structured policy with graph neural networks

    Tingwu Wang et al. “Nervenet: Learning structured policy with graph neural networks”. In:International conference on learning representations. 2018

    2018

  26. [34]

    D(R,O)Grasp: A Unified Representa- tion of Robot and Object Interaction for Cross-Embodiment Dexterous Grasping

    Zhenyu Wei et al. “D(R,O)Grasp: A Unified Representa- tion of Robot and Object Interaction for Cross-Embodiment Dexterous Grasping”. In:2025 IEEE International Confer- ence on Robotics and Automation (ICRA). 2025, pp. 4982– 4988.DOI:10.1109/ICRA55743.2025.11127754

    2025

  27. [35]

    Biomechanics and Motor Control of Human Movement

    DA Winter. “Biomechanics and Motor Control of Human Movement”. In: ed. by DA Winter et al. 2nd. New York: John Wiley & Sons, 1990. Chap. 10, pp. 250–280

    1990

  28. [36]

    DensePhysNet: Learning Dense Physical Object Representations Via Multi-Step Dynamic Interac- tions

    Zhenjia Xu et al. “DensePhysNet: Learning Dense Physical Object Representations Via Multi-Step Dynamic Interac- tions”. In:Robotics: Science and Systems. 2019

    2019

  29. [37]

    A Learning-Based Quadcopter Con- troller With Extreme Adaptation

    Dingqi Zhang et al. “A Learning-Based Quadcopter Con- troller With Extreme Adaptation”. In:IEEE Transactions on Robotics41 (2025), pp. 3948–3964.DOI:10.1109/TRO. 2025.3577037

    2025

  30. [38]

    Robot Parkour Learning

    Ziwen Zhuang et al. “Robot Parkour Learning”. In:Confer- ence on robot learning. 2023

    2023

Pith tools

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