Pith. sign in

REVIEW 2 major objections 5 minor 38 references

Rapid Embodiment Adaptation for Quadrupedal Locomotion

T0 review · 2 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 The paper establishes that a quadruped can infer its own current joint-limit and mass parameters from a 0.4-second window of actions and observations, and that conditioning a generalist locomotion policy on this inferred embodiment keeps…

desk verdict A solid, incremental online embodiment-identification result with real hardware gains, but the abstract overstates the success rates: under the hardest conditions the robot actually succeeds in only 30–50% of trials, and the 'locked leg' is a 0.1-scale virtual limit, not a hard lock. read the letter →

arxiv 2608.01506 v2 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 quadrupedallocomotionembodimentadaptationonlinesystemidentificationcross-embodimentpolicyjoint-limitconstraintstrunkmasspayloadreinforcementlearningLSTMencoder
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 a robot can treat its own body as an unknown to be re-identified online, rather than as a fixed, known specification. It proposes a two-stage setup: a generalist locomotion policy trained under randomized joint-limit and trunk-mass changes, plus a lightweight adaptation module that reads a short history of proprioceptive observations and actions and predicts the current joint-range scaling and trunk mass. In simulation the estimated parameters converge within about 0.4 seconds, and the closed-loop controller outperforms a policy that adapts implicitly from history. On a real Unitree Go2, the system keeps walking under a fully locked front leg and a 5 kg payload, conditions where the non-adaptive baseline falls. If the result holds, embodiment-conditioned policies no longer need a static, pre-known hardware description at deployment.

What carries the argument

The load-bearing object is the embodiment description $\hat{\phi}_t = (\{\hat{d}_{t,j}\}, \hat{m}_t)$, where each per-joint adaptable descriptor is $\hat{d}_{t,j} = (\hat{q}^0_{t,j}, \hat{\tau}^{\max}_{t,j}, \hat{q}^{\min}_{t,j}, \hat{q}^{\max}_{t,j})$ and $\hat{m}_t$ is the trunk-mass offset, together predicted by an adaptation module $f(h_t)$ from a sliding 20-step history of observations and actions. The module is an LSTM over per-joint history features followed by a Transformer that attends across joints, trained by supervised regression to the ground-truth parameters recorded in offline simulation trajectories. A latent variant instead predicts the frozen policy's embedding $z_t = E_{\mathrm{lat}}(d_t)$ under the same objective. The mechanism that carries the argument is the two-stage decomposition: oracle-conditioned reinforcement learning produces a policy that already knows how to move under each embodiment, and the supervised estimator supplies the missing condition at run time.

What would settle it

Take the same real robot and replace the virtual joint lock with a mechanical stop that physically caps the front-right leg's range, then run the estimator and measure whether predictions still converge to the true range within 0.4 seconds and whether the robot still walks the distances reported in Table IV; if it falls within the non-adaptive failure time or the estimate drifts, the transfer depends on the virtual-limit equivalence rather than on genuine joint identification.

Watch

Extended reading notes

Core claim

The central claim is that explicit online embodiment identification makes cross-embodiment locomotion policies practical under changing hardware. The authors train a cross-embodiment base policy with reinforcement learning, conditioning it on ground-truth embodiment descriptors that include per-joint nominal position, torque limit, and joint-range endpoints plus trunk mass, while randomizing these quantities at training time. They then freeze that policy and train an adaptation module by supervised learning to map a sliding window of 20 past observation-action pairs to either the explicit physical descriptors or a latent embedding of them. At deployment the policy conditions on the module's estimate each timestep, forming a closed loop that converges to the new embodiment within roughly 0.4 seconds after an abrupt change. The paper reports that in simulation this closes most of the gap to an oracle policy, and on hardware it sustains stable locomotion under a front-right leg whose joint range is scaled to 0.1 and under a 5 kg payload, where the no-adaptation baseline fails early.

Load-bearing premise

The load-bearing premise is that the real-world way of restricting a joint—a high-gain controller that holds the joint outside a prescribed range—produces the same interaction dynamics as the hard kinematic joint limits used in simulation, so the estimator trained on simulated stops transfers to the hardware's virtual stops.

Editorial extensions

If this is right

  • Within about 0.4 seconds of an abrupt change, the estimator converges to the new joint-range scale or trunk mass, so adaptation happens mid-episode without retraining or reset.
  • Conditioning on the inferred embodiment lets the base policy switch gaits—an asymmetric crutch-like gait for a locked leg, a load-bearing stance for a heavy payload—rather than falling.
  • In simulation, both explicit and latent representations substantially outperform the implicit end-to-end policy that maps history directly to actions, suggesting that an explicit parameter bottleneck is not a handicap.
  • For the hardest real-world settings (joint range scaled to 0.1 and 5 kg payload), the non-adaptive policy falls early while the adapted robot continues, so the method extends embodiment-conditioned policies to unknown deployment-time hardware.

Reading between the lines

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

  • Because the explicit descriptor already includes torque limits, nominal position, and velocity limits, the same estimator architecture could be pointed at other per-joint degradations without retraining the base policy; the paper only varies joint ranges and trunk mass.
  • The 0.4-second convergence at 50 Hz implies the module could track slowly drifting embodiment changes—battery drain, actuator heating, terrain-influenced effective mass—as well as abrupt ones, though the paper does not test drift.
  • The real-world joint lock is enforced by a high-gain controller outside a prescribed range, not a mechanical stop; a natural test is whether a truly jammed actuator produces the same interaction signature the estimator was trained on.
  • If explicit identification scales to more parameters, it could support a robot reporting its own hardware state—for example, "front-right knee limited to 30 percent range"—during deployment, enabling operator diagnosis and safety monitoring.
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

2 major / 5 minor

Summary. The paper proposes a two-stage framework for online embodiment adaptation in quadrupedal locomotion. It first trains a URMA-based cross-embodiment policy with embodiment randomization in simulation, then trains a lightweight LSTM-plus-Transformer adaptation module to infer either explicit physical parameters (joint-limit scaling, torque limits, nominal position, trunk mass offset) or latent embodiment embeddings from a 20-timestep interaction history. The method is evaluated in simulation for identification accuracy and closed-loop command tracking, and on a real Unitree Go2 under static and mid-episode joint-range constraints and payload additions, with comparisons to a no-adaptation baseline and an implicit end-to-end policy.

Significance. If the quantitative claims are suitably corrected, this is a useful contribution: the two-stage decomposition decouples expensive policy training from lightweight online identification, the explicit representation yields inspectable physical estimates, and the real-hardware trials show consistent improvements over a no-adaptation baseline across all reported conditions. The simulation ablations over architecture and history length, the closed-loop comparison across five methods, and the inclusion of both static and abrupt mid-episode embodiment changes are strengths. The central limitation is that the headline wording overstates the hardware results: the two most severe conditions show only 30% and 50% raw success rates, and the most constrained joint condition is a 0.1 scale, not a fully locked joint.

major comments (2)
  1. [V-C, Table IV vs. Abstract and Discussion] The headline claim that the robot 'maintains stable locomotion' under a fully locked leg and a 5 kg payload is not supported by the paper's own quantitative data. The table note states that a normalized success score of 1.0 corresponds to an 80% raw success rate; therefore Ours at FR Joint Lock 0.1 (normalized 0.375) corresponds to 30% raw success, and Ours at Payload 5.0 kg (normalized 0.625) corresponds to 50% raw success over 10 trials. In these two headline conditions the robot falls or exceeds the joint-velocity limit in 50–70% of trials. Moreover, 'fully locked leg' is not actually tested: the most severe joint condition is a joint-limit scale of 0.1, i.e., 10% of the nominal range, not 0.0. Please report raw trial counts and revise the abstract and Discussion to claim consistent improvement, longer survival time, and greater distance traveled, rather than stable locomotion in an absolute sense.
  2. [IV.A and V.C] The simulated joint-limit model and the real-hardware joint-limit implementation are not shown to be equivalent, and this mismatch is load-bearing for the 'locked leg' claim. In simulation, the policy and estimator are trained with kinematic joint-limit bounds from the randomization in Table II, while in hardware the constraint is a virtual limit implemented by 'applying a high control gain outside a prescribed joint range' (Section IV.A). A high-gain spring permits overshoot and changes the effective joint impedance, so it is not identical to the hard kinematic stop used during training. The paper provides no evidence, such as realized joint-range traces or torque-position comparisons, that the virtual limit approximates the simulated hard limit. Without such a validation, the transfer of the learned estimator to genuinely locked or severely constrained joints is questionable; please add a quantitative equivalence check or soften the claim to describe the tested virtual-limit conditions.
minor comments (5)
  1. [III.C.3, Eq. (1)] The supervised objective is stated ambiguously for the latent variant: f_lat outputs (z_hat, m_hat), but the target is written as rho_t = xi_t = (d_fix, z_t, m_t). Since d_fix is fixed and not predicted, the loss as printed cannot be computed from f_lat's outputs. Please state the explicit and latent losses separately with the exact target variables.
  2. [Fig. 4] The y-axis label 'RMSE (x10^2)' appears to be a factor-of-ten error: the plotted values are around 1–7, which would be consistent with an RMSE on the order of 10^-2 as in Table III, not 10^2. Please correct the label.
  3. [Table III] The table heading calls the metric 'Error rate', but the values are RMSE. Please specify the units and state whether the RMSE is averaged over all predicted joint and body parameters.
  4. [V.C (dynamic embodiment changes)] The mid-episode change experiments are reported only qualitatively with representative trials. Given that the static conditions already show 30–50% success in the hardest cases, please report quantitative success rates or trial counts for the abrupt-change conditions as well.
  5. [Throughout] There are minor typographical and reference issues: Section III.C contains 'paramters'; Section IV.A ends with a double period; and references [22] and [23] duplicate the same GET-Zero paper. Please clean these up.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the explicit estimator is trained on simulator ground-truth embodiment parameters and the central hardware claim is benchmarked against a no-adaptation baseline; only the non-central latent variant is self-referential.

full rationale

The central explicit-representation pipeline is not circular. The cross-embodiment base policy is trained with oracle embodiment descriptions (Section III-B), and the adaptation module f_exp is trained by supervised regression to the simulator's ground-truth per-joint limits, torque limits, and trunk mass (Sections III-C.3, IV-C); the supervision signal is external to the policy and not derived from the predicted outcome. The real-world evaluation is an external, falsifiable comparison against a No Adaptation baseline on a physical Unitree Go2, with failures defined by falls or joint-velocity exceedance (Section V-C). The S_rel normalization in Eq. 2 is explicitly labeled 'by construction' and is only a reporting metric, not a prediction. The latent variant does contain a self-referential element: its supervision target is the frozen base-policy encoder output z_t = E_lat(d_t) (Section III-C.1), so latent adaptation success measures distillability of the policy's own embedding rather than identification of external physical quantities. This is not load-bearing because the paper selects and validates the explicit representation on hardware (Section V-C) and makes no physical-identification claim for the latent variant. Separately, Eq. 1 as printed uses hat-variables (rho_t = hat_phi, rho_t = hat_xi_t) as targets, which if read literally would make the loss trivial; the surrounding text and nonzero RMSE results (Table III) show the intended target is ground truth, so this is a notation slip rather than a circular derivation. The abstract's 'fully locked leg' overstates Table IV's 0.1 joint-limit condition (30% success), but that is an evidence-claim mismatch, not circularity.

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

The central claim rests on simulator fidelity, on the equivalence between the real robot's virtual joint limit and the simulation's kinematic limit, and on the robustness of the frozen policy to noisy estimates. There are no invented physical entities; the embodiment parameters are supervised targets from the simulator.

free parameters (3)
  • History window length T = 20 timesteps (0.4 s at 50 Hz)
    Chosen from Fig. 4 as a practical trade-off between accuracy and latency; the headline 0.4 s convergence follows from this choice.
  • Real-world failure threshold = joint velocity > 15 rad/s terminates a trial
    Hand-set criterion defining success in Table IV; changing it would change reported success rates.
  • Joint-limit randomization range = [0, 0.5] scale during training
    Defines the distribution the base policy and estimator were trained on; simulation extrapolates to 1.0 and real hardware tests 0.1 and 0.3.
assumptions (3)
  • domain assumption The IsaacLab simulator reproduces Unitree Go2 dynamics and the embodiment variations used in training.
    All policy and estimator training and most closed-loop evaluation occur in simulation; real-hardware transfer depends on this fidelity (Sections IV and V).
  • domain assumption A high control gain outside a prescribed joint range on the real robot behaves like the kinematic joint limits used in simulation.
    Section IV.A describes the real constraint implementation; the paper does not measure whether the resulting dynamics match simulated hard limits.
  • domain assumption The frozen URMA base policy remains effective when conditioned on noisy online estimates produced by the adaptation module.
    The two-stage training does not guarantee robustness to estimation error; sim and real results are the only evidence (Section III.C, V).

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…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 7 linked inside Pith

  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)

    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)

    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)

    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)

    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

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