Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Towards bridging the gap: Systematic sim-to-real transfer for diverse legged robots

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read About 20 seconds of encoder-only in-air data are enough to train legged-robot controllers that transfer zero-shot, without dynamics randomization.

desk verdict Solid engineering with real hardware validation, but the headline energy claim is a cross-paper comparison and the in-air-to-stance parameter assumption is asserted, not tested. read the letter →

arxiv 2509.06342 v2 pith:JYTU36PO submitted 2025-09-08 cs.RO

Filip Bjelonic , Fabian Tischhauser , Marco Hutter This is my paper · ORCID
classification cs.RO
keywords leggedrobotsquadrupedallocomotionsim-to-realtransferreinforcementlearningenergyefficiencysystemidentificationcostoftransportactuatordynamics
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 the simulation-to-reality gap for legged robots is mostly an actuator-dynamics gap, and that a deliberately small set of physically meaningful parameters can close it. Its central claim is that fitting four per-joint quantities plus one global delay, a total of $4n+1$ numbers, from about twenty seconds of in-air and encoder-only motion makes a simulator accurate enough that blind (proprioception-only) locomotion policies transfer to hardware in one shot, with no dynamics randomization. If true, this would make sim-to-real deployment cheaper and more accessible: no torque sensors, minutes of data, or large hand-tuned reward stacks. The paper also reports that the fitted simulator enables energy-efficient gaits, with a 32 percent reduction in ANYmal's full cost of transport to 1.27.

What carries the argument

The engine of the argument is the per-joint closed-loop model in Eq. (6): each actuator is treated as a constant-inertia, constant-damping second-order system driven by saturated PD torque, with an added joint-position bias and a global command delay. The parameter vector $p=(I_a, d, \tau_f, \tilde{q}_b, T_d) \in \mathbb{R}^{4n+1}$ is fitted by minimizing the time-averaged squared joint-position error between real and simulated chirp replays, using CMA-ES over thousands of parallel environments. Once fitted, this model carries the sim-to-real transfer: the policy is trained in the fitted simulator with a compact reward whose energetic term is the PMSM electrical-plus-mechanical loss model, and it is deployed without dynamics randomization.

What would settle it

Fit the 4n+1 parameters from in-air chirps, then stand the robot on a force plate and inject a known PD torque perturbation during stance; if the simulated joint response diverges from the measured one as posture changes while the in-air fit is held fixed, the claim that in-air data suffices for contact tasks fails. A second decisive check is to train the PACE policy with no dynamics randomization and find any terrain or speed at which it fails while a randomization-trained policy succeeds.

Watch

Extended reading notes

Core claim

The paper's central discovery is a bottom-up identification recipe, called PACE: suspend the robot, drive all joints with chirp signals, replay the commands in a massively parallel simulator, and use evolutionary search to fit per-joint armature inertia, viscous damping, Coulomb friction, and joint bias, together with one global command delay. The fitted simulator reproduces in-air joint trajectories nearly exactly, generalizes to unseen PD gains and trajectories, and then serves as the training environment for a blind policy with a four-term reward built around a physics-grounded PMSM loss model. Deployed zero-shot on hardware, the policy walks, runs, balances on two legs, and climbs stairs; on ANYmal it lowers the cost of transport from 1.86 to 1.27 and extends a full-charge run from 2.64 km to 4.12 km. The same parameterization is reported to fit both open and closed platforms.

Load-bearing premise

The load-bearing premise is that a constant inertia and damping measured while the legs swing freely in air remain the right joint dynamics when the feet are on the ground, even though the effective inertia at each joint is then dominated by the body mass and varies with posture.

Editorial extensions

If this is right

  • Robots without joint-torque sensors can obtain deployable locomotion policies from about 20 seconds of encoder-only, in-air data, removing a practical barrier for small or closed platforms.
  • Because dynamics randomization is no longer needed, policies trained in the fitted simulator can exploit the machine's actual dynamics, producing the reported straighter-knee gaits and the ANYmal cost-of-transport drop to 1.27.
  • The same 4n+1 parameterization is reported to transfer across platforms, from a 52 kg quadruped to a 4 kg largely 3D-printed one, suggesting the identification is not tied to one drive design.
  • With the model fitted, the remaining ceiling on running speed is hardware, the battery current limit, rather than model error, which shifts the practical bottleneck to power electronics and sensing.

Reading between the lines

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

  • The paper leaves implicit that the constant-parameter model may saturate at higher step frequencies or on strongly varying terrain, since the effective inertia at each joint changes with stance posture; a testable extension is to add a posture-dependent inertia term and see whether it widens the no-randomization envelope.
  • The identification recipe is not obviously limited to quadrupeds: any articulated robot with PMSM drives and encoder-only joint sensing could in principle be fit from fixed-base chirps, so the claim suggests a general sim-to-real shortcut for arms and humanoids.
  • A head-to-head study of the same policy with and without dynamics randomization on identical terrain could separate how much of the 32 percent cost-of-transport gain comes from removing randomization versus from the energy reward.
  • Temperature and wear will shift the fitted parameters over time, so a lightweight online update of the global scale of the inertia and damping terms is a natural follow-up that the paper names as future work.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 PACE, a sim-to-real framework for legged robots. The method identifies a compact per-joint parameter set (armature inertia, viscous damping, Coulomb friction, joint bias, and a global delay; 4n+1 parameters total) by fitting simulated fixed-base, in-air joint trajectories to real encoder-only chirp data using CMA-ES in massively parallel simulation. Policies are then trained with PPO in the fitted simulator without dynamics randomization, using a four-term reward that includes a physics-based electrical and mechanical loss model. The authors validate the pipeline bottom-up on a single drive, on full robots in air, and in on-ground locomotion on ANYmal, TYTAN, and MINIMAL, with additional deployments on ten other robots. They report a 32% reduction in ANYmal's full Cost of Transport to 1.27, and a TYTAN CoT of 0.97, measured on a 400 m running track via battery state of charge.

Significance. If the central claims hold, the paper would make a valuable practical contribution: roughly 20 seconds of encoder-only in-air data per robot would suffice to train deployable, blind locomotion policies without dynamics randomization, while also improving energy efficiency through a first-principle energy reward. The bottom-up validation is a genuine strength: the single-drive experiments recover analytic inertia values within 2--15%, the full-robot in-air comparison shows clear improvement over a URDF-only baseline and competitive behavior against a learned actuator network, and the cross-platform deployment across three primary and ten additional robots is impressive. The promise of open-sourced code, models, and data would enhance reproducibility. However, the two headline claims -- that in-air identification transfers to stance without randomization, and that the 32% CoT reduction is attributable to the proposed method -- are not yet established by the experiments as presented.

major comments (3)
  1. [Section 5, Section 2.2, Appendix B] The central takeaway (Section 6) asserts that constant per-joint parameters {Ia, d, tau_f, q_tilde_b, Td} identified from fixed-base, zero-contact in-air data are sufficient to eliminate dynamics randomization in stance. Section 2.2 itself concedes that load-independent damping is 'common but imperfect' and that torque--current maps can be state dependent, and Appendix B (Fig. 23c) shows that the effective inertia at the HFE joint during horizontal on-ground motion reaches roughly 8--10 kg m^2, orders of magnitude above the in-air values. Section 4.1.2 further reports a full-robot Ia about four times the single-drive value, indicating that the fitted inertia absorbs configuration-dependent link and compensation effects. The manuscript offers no experiment that re-identifies the parameters under load, compares simulated stance torques against measured torques on a torque-instrumented platform, or otherwise tests whether the load-independence assumption holds during contact. Since PACE deliberately omits dynamics randomization as a safety net, this extrapolation is load-bearing and needs direct evidence rather than the single-sentence assertion 'In-air data suffices for contact tasks.'
  2. [Section 3.3.3, Table 3, Abstract] The headline 32% Cost of Transport reduction compares ANYmal D (CoT 1.27) with ANYmal C (CoT 1.86) from the previously published reference [71], not with a same-hardware baseline. Table 3 shows that the two runs also differ in ambient temperature (26 deg C vs 10 deg C), and ANYmal C and ANYmal D may differ in actuators, firmware, mass, and controller generation. The entire reduction is attributed to PACE without controlling for these factors. To support the abstract's claim, the authors need a controlled comparison on the same robot unit, for example running the actuator-network baseline or a domain-randomized baseline on the same ANYmal D with the same reward and track protocol, or running PACE on ANYmal C, or otherwise demonstrating that hardware/firmware/temperature differences do not explain the CoT gap.
  3. [Section 2.3.3, Table 4] The paper claims the physics-grounded energy reward contributes to the reported efficiency gains, but no ablation isolates the effect of the energy term. The reward scales and scheduling (c_v, c_e, c_ftd, c_c, lambda half-life) are free parameters listed in Eq. (20) and Table 2, and P_pot in Eq. (14) averages to zero over flat cyclic locomotion, so it is unclear how much of the improvement comes from the energetic reward versus the fitted simulator enabling different gaits. Since one of the four stated contributions is the energetic assessment, an ablation on at least one platform -- same fitted simulator with and without the energy term, or with a torque-squared proxy -- measured by battery CoT would quantify this contribution.
minor comments (5)
  1. [Abstract and Data Accessibility Statement] The abstract states that all code, models, and datasets are publicly available, while the Data Accessibility Statement says the source code 'will be released soon' and the text elsewhere says 'will be released'; please align these statements.
  2. [Table 5] The header for the armature row reads 'Armature I_a [10^-3 kg m^3]'; the unit should be kg m^2, not kg m^3.
  3. [Section 6 and Section 4.2.4] The text refers to 'current limits (Section 12)', which appears to be an incorrect internal reference; it should point to Figure 12 or Section 4.2.4 where the 32 A battery limit is described.
  4. [Section 5] In the discussion of knee extension, the text says 'Table 4.2.3'; this should be 'Table 4'.
  5. [Appendix B, Eq. (42)] Equation (42) appears to have a typo: the inertia term should contain sin^2(q_k/2), not sin(q_k/2), based on the derivation from Eqs. (34)--(40).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the in-air parameter fit is an input to the simulator, while the headline transfer and CoT results are measured on hardware from battery state-of-charge and state estimation, not reconstructed from the fit.

full rationale

The paper's central derivation is not circular. The parameter vector p = [Ia, d, tau_f, q_tilde_b, Td] is optimized against an explicit in-air trajectory-matching objective (Eqs. 3-4), and the fitted quantities are joint-position tracking errors, not the deployment metrics (base-velocity tracking, zero-shot locomotion success, or Cost of Transport). The CoT reduction to 1.27 is measured from battery state-of-charge over a 400 m track (Eqs. 25-27), using hardware constants R, k_i, r in the energy model (Eq. 12); it is not derived from the fitted parameters by construction. The energy reward penalizes a physical loss model and is not fitted to the reported CoT. The paper's heavy self-citation (refs 20, 32, 45, 71) supplies baselines, RL infrastructure, reward templates, and a track protocol, but none of these is a load-bearing uniqueness theorem or an unverified ansatz smuggled in by citation; the core claim is supported by direct hardware deployment across three platforms and ten additional systems. The skeptic's concern that actuator parameters fitted in air may not remain valid under stance loads (Appendix B, Fig. 23) is an extrapolation-risk/correctness criticism, not a circularity: the ground behavior is not equivalent to the in-air fit by construction, and the paper explicitly lists load-independence as an empirically-satisfied assumption rather than as a definitional identity. Eq. (7) is used to argue for fixing PD gains during identification, which is an identifiability precaution, not a circular reduction. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

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

The ledger captures what PACE assumes rather than derives: the joint model structure (Eq. 6), the transferability of in-air fits to stance, the PMSM loss model, and several hardware linearity assumptions. No genuinely new physical entities are introduced; the 'virtual inertia' observed with firmware compensation is an empirical lumped effect, not a new physical object.

free parameters (6)
  • Per-joint armature inertia I_a = Table 5, e.g. ANYmal 39-89 x10^-3 kg m^2, TYTAN HFE 106-120 x10^-3 kg m^2
    Fitted via CMA-ES (Eq. 4) to in-air chirp replay; central to simulator alignment.
  • Per-joint viscous damping d = Table 5, e.g. ANYmal 4.3-5.5, TYTAN 0.17-3.9 N m s rad^-1
    Fitted with the same trajectory-matching objective; contributes to transfer fidelity and the energy reward.
  • Per-joint Coulomb friction tau_f = Table 5, e.g. ANYmal 0.0029-0.045 in normalized units
    Fitted; the paper does not report SI units for friction (Table 5 header 'Friction tau_f [-]').
  • Per-joint joint position bias q_bias = Table 5, values in radians around +/-0.03
    Fitted to absorb encoder offsets and zero-position errors.
  • Global command delay T_d = 7.5 ms for TYTAN and ANYmal; 0.0 for MINIMAL
    Fitted; lumps communication and control delays in the joint loop.
  • Reward scales and scheduling (c_v, c_e, c_ftd, c_c, lambda half-life) = Table 2: c_v=0.2, c_e=-16e-5 (TYTAN/ANYmal) or -128e-5 (MINIMAL), c_c=-1.0, c_ftd=-0.1; half-life 500 iterations
    Hand-tuned; they shape the learned gait and the efficiency outcome and are not identified from data.
assumptions (6)
  • domain assumption The per-joint closed loop can be modeled as Ia*qdd + d*qd = sat(P*(qhat - q + qb) - D*qd + tau_comp) + tau_f (Eq. 6)
    This second-order plus saturation model is the core of the identification; the paper acknowledges load-independent damping is 'common but imperfect' (Section 2.2).
  • domain assumption In-air fixed-base identification transfers to on-ground stance despite base-inertia dominance
    Stated in Section 5; Appendix B shows stance inertia varies by 1-2 orders of magnitude and is not captured by the constant-parameter model, so transfer is an empirical assertion.
  • domain assumption High-bandwidth current control or LTI-approximable drive with the inner torque loop as near-unit gain
    Listed in Section 2.4 assumptions; verified only for TYTAN (Section 3.2.1, 346 Hz bandwidth) and assumed for ANYmal and MINIMAL.
  • domain assumption Electrical losses dominated by q-axis current with negligible i_d and constant R, k_i, gear ratio (Eq. 12)
    Used to define the energy reward; Table 1 provides hardware constants, but the q-axis dominance is an idealization.
  • domain assumption Mild temperature dependence and sufficient structural stiffness during data collection
    Explicit assumptions in Section 2.4; the paper later notes temperature and aging shift effective parameters as a limitation.
  • domain assumption Correct kinematics (URDF/USD, frames) and an adequate rigid-body model in the simulator
    Stated in Section 2.4 assumptions; if kinematics are wrong, the identified parameters absorb the error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards bridging the gap: Systematic sim-to-real transfer for diverse legged robots." pith.science (2026). https://pith.science/paper/JYTU36PO

@misc{pith2026250906342,
  author       = {Pith},
  title        = {Pith review of: Towards bridging the gap: Systematic sim-to-real transfer for diverse legged robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYTU36PO}},
  note         = {Machine review of arXiv:2509.06342}
}
read the original abstract

Legged robots must achieve both robust locomotion and energy efficiency to be practical in real-world environments. Yet controllers trained in simulation often fail to transfer reliably, and most existing approaches neglect actuator-specific energy losses or depend on complex, hand-tuned reward formulations. We propose a framework that integrates sim-to-real reinforcement learning with a physics-grounded energy model for permanent magnet synchronous motors. The framework requires a minimal parameter set to capture the simulation-to-reality gap and employs a compact four-term reward with a first-principle-based energetic loss formulation that balances electrical and mechanical dissipation. We evaluate and validate the approach through a bottom-up dynamic parameter identification study, spanning actuators, full-robot in-air trajectories and on-ground locomotion. The framework is tested on three primary platforms and deployed on ten additional robots, demonstrating reliable policy transfer without randomization of dynamic parameters. Our method improves energetic efficiency over state-of-the-art methods, achieving a 32 percent reduction in the full Cost of Transport of ANYmal (value 1.27). All code, models, and datasets are publicly available.

Figures

Figures reproduced from arXiv: 2509.06342 by the authors.

Figure 1
Figure 1. Comparison of real and simulated robot trajectories under a 0.1 – 2.0 Hz chirp input. (a) Picture of the real robot with an overlaid trace showing its motion sequence from experiment. (b) Equivalent view in simulation using our modeling approach. In both (a) and (b), the robot’s base is suspended so that its legs swing freely. (c) Phase portrait of the “delta” joint state—i.e., at each timestep, the real joint posit… view at source ↗
Figure 2
Figure 2. Representative legged robots with explicitly documented use of PMSM for actuation. actuator-aware formulations (52, 53), we propose a loss model tailored to PMSMs that captures the dominant sources of energy dissipation while remaining tractable within modern reinforcement learning simulators. This balance enables training of locomotion policies that are explicitly energy-efficient. 1.4 Notation We adopt the followi… view at source ↗
Figure 3
Figure 3. Overview of the proposed Precise Adaptation through Continuous Evolution (PACE) pipeline with policy training. (i) Collection of real in-air data on a fixed-base setup (top left). (ii) Evolutionary parameter fitting of joint dynamics to align simulated and measured trajectories (top right). (iii) Blind policy training in simulation with zero-shot deployment on hardware (bottom). replay the recorded joint targets at … view at source ↗
Figures from the paper (18 more)
Figure 4
Figure 4. Figure 4: Primary robotic platforms evaluated in this study. Top row (left to right): ANYMAL, TYTAN, and MINIMAL. Additional systems include Sony’s AIBO, Softbank’s NAO, ALMA, SPACEHOPPER, LEVA, MAGNECKO v2 and v1, and Fourier’s GR-1. All robots are depicted in operation using t…
Figure 5
Figure 5. Figure 5: Experimental setups and representative evaluation across abstraction levels. of this mass, the joint inertia can be adjusted in the range 1.65 – 6.27 × 10−2 kg m2 . Current loop validation. The goal of this experiment is to verify the motor drives’ ability to precisely…
Figure 6
Figure 6. Figure 6: Sim-to-real evaluation of TYTAN using the proposed PACE framework. Prepared using TRR.cls [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Sim-to-real evaluation on ANYMAL. Panels (a–c) illustrate real-world deployments. Panel (d) shows the corresponding phase portraits. Panel (e) presents commanded and measured forward velocities (left) and steady-state velocity error distributions across methods (right)…
Figure 8
Figure 8. Figure 8: Running-track evaluations across robotic platforms. (a) Representative stills from one step cycle. (b) Battery consumption over distance (SoC). (c) Cost of transport (CoT) with contributions from electronics (CoE, diagonal hatching) and drives’ inverter switching (CoD,…
Figure 10
Figure 10. Figure 10: In-air LF–HFE joint trajectory of ANYMAL (zoomed view). Dashed gray: target trajectory. Dashed green: measured trajectory. Blue: no actuator model. Red: actuator network baseline. Orange: proposed method (PACE). torques on the LF leg also match well; small deviations …
Figure 11
Figure 11. Figure 11: Dynamic-limit demonstrations. Top: two-legged balance with ANYMAL. Center: running with ANYMAL. Bottom: stair climbing with MINIMAL [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Running experiments with ANYMAL. Top: commanded forward velocity versus measured base velocity (state estimation). Bottom: battery current profile during the same run, including measured current and the saturated peak limit of 32 A imposed by the battery. Prepared usi…
Figure 13
Figure 13. Figure 13: Idealized PMSM saturation envelope for the TYTAN hip motor, showing motor velocity ω (x-axis) versus torque τ (y-axis). The solid blue curve indicates the enforced torque limits in simulation: commanded torques inside the envelope are applied directly, while those out…
Figure 14
Figure 14. Figure 14: Simplified block diagram of the motor-driven joint actuator, illustrating the electronics (left, yellow) and rigid-body dynamics (right, red) subsystems. In the electronics path, a PD position controller with gains Pτ and Dτ drives a PI current loop—comprised of integ…
Figure 15
Figure 15. Figure 15: Validation of TYTAN on the HFE joint using proportional–derivative (PD) gains of 145/5. Shown are commanded trajectories, real measurements and the response under the near-optimal identified model. −20 0 Magnitude |Hi| [dB] 100 101 102 Frequency f [Hz] −150 −100 −50 0…
Figure 18
Figure 18. Figure 18: An example bode plot from the single-drive experiment for the closed-loop transfer function Hq : ˆqj → qj , mapping commanded joint position qˆj to measured joint position qj . Because the dependencies are strongly nonlinear, their derivatives are large, and the effec…
Figure 19
Figure 19. Figure 19: An example bode plot from the single-drive experiment for the closed-loop transfer function Hq : ˆqj → qj , mapping commanded joint position qˆj to measured joint position qj . The hatched region indicates the 2σ confidence interval from analytic estimation (Appendix …
Figure 20
Figure 20. Figure 20: Minimal trajectories, LF KFE C Link Inertia Estimation To analyze the discrepancy between single-drive and full￾robot inertia estimates, we performed a study on a leg￾segment of TYTAN. The KFE was disconnected to minimize friction, allowing the thigh or shank to swing…
Figure 21
Figure 21. Figure 21: Additional sim-to-real evaluation of ANYMAL. (a) Sideways velocity tracking. (b) Yaw angular velocity tracking. Both panels compare commanded and measured base velocities over time (left) and summarize steady-state errors across methods (right). These results explain …
Figure 22
Figure 22. Figure 22: Joule heating constants of different robots, derived from motor parameters over body weight. These values offer a first-order estimate for black-box robotic systems. Prepared using TRR.cls [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]
Figure 24
Figure 24. Figure 24: Breakdown of reduced inertia contributions for the full robot configuration for TYTAN in Section 3.2.2. The bar plot shows the distribution of components shaping the fitted joint inertia estimate Ia (dotted black line) obtained with PACE. The solid line denotes the ±1…
Figure 25
Figure 25. Figure 25: Full battery depletion time of ANYMAL (Top) and TYTAN (bottom) at walking in blue, resting (Electronics & drives) in orange, and lying on the ground in green (Electronics). Prepared using TRR.cls [PITH_FULL_IMAGE:figures/full_fig_p026_25.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Mask2Real-WM: Segmentation Masks as a Sim-to-Real Bridge for Controllable Dexterous World Models

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Segmentation-space dynamics pretrained on 50+ hours of simulation, then fine-tuned on under 2.5 hours of real data, plus a ControlNet RGB renderer, give per-DoF controllability across a 23-DoF dexterous hand.

  2. Actuator Reality Shaping for Zero-Shot Sim-to-Real Robot Learning

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A per-joint 2-DoF feedforward–feedback controller with disturbance observer shapes real actuators to match idealized second-order sim dynamics, enabling zero-shot RL policy transfer.

  3. Efficient Real-World Online Reinforcement Learning for Robot Manipulation via Centralized Training and Critic Decomposition

    cs.RO 2026-08 conditional novelty 4.0 of 10

    A reinforcement learning framework that shares a decomposed critic between arm and gripper policies improves real-world pick-and-place success rates by 20 to 70 percentage points over a strong baseline.

Reference graph

Works this paper leans on

73 extracted references · 55 canonical work pages · cited by 3 Pith papers

  1. [71]

    Bjelonic, F., J. Lee, P. Arm, D. Sako, D. Tateo, J. Peters, and M. Hutter. Learning-based design and control for quadrupedal robots with parallel-elastic actuators.IEEE Robotics and Automation Letters, V ol. 8, No. 3, 2023, pp. 1611–1618

    work page 2023

  2. [1]

    Gehring, A

    Hutter, M., C. Gehring, A. Lauber, F. Gunther, C. D. Bellicoso, V . Tsounis, P. Fankhauser, R. Diethelm, S. Bachmann, M. Bl ¨osch, et al. Anymal-toward legged robots for harsh environments.Advanced Robotics, V ol. 31, No. 17, 2017, pp. 918–931

    work page 2017

  3. [2]

    Motor SDK Development Guide

    Unitree Robotics. Motor SDK Development Guide. Online:https://support.unitree.com/home/en/ Motor_SDK_Dev_Guide/overview, 2025. Accessed: Aug. 21, 2025

    work page 2025

  4. [3]

    L ´eziart, T

    Aractingi, M., P.-A. L ´eziart, T. Flayols, J. Perez, T. Silander, and P. Sou `eres. Controlling the solo12 quadruped robot with deep reinforcement learning.scientific Reports, V ol. 13, No. 1, 2023, p. 11945

    work page 2023

  5. [4]

    G.A low cost modular actuator for dynamic robots

    Katz, B. G.A low cost modular actuator for dynamic robots. Ph.D. thesis, Massachusetts Institute of Technology, 2018

    work page 2018

  6. [5]

    Liu, D., F. Yang, X. Liao, and X. Lyu. Diablo: A 6-dof wheeled bipedal robot composed entirely of direct-drive joints. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 3605–3612

    work page 2024

  7. [6]

    Haemers, J

    Derammelaere, S., M. Haemers, J. De Viaene, F. Verbelen, and K. Stockman. A quantitative comparison between BLDC, PMSM, brushed DC and stepping motor technologies. In 2016 19th International Conference on Electrical Machines and Systems (ICEMS). Ieee, 2016, pp. 1–5

    work page 2016

  8. [7]

    Gamazo-Real, J. C., E. V ´azquez-S´anchez, and J. G ´omez-Gil. Position and speed control of brushless DC motors using sensorless techniques and application trends.sensors, V ol. 10, No. 7, 2010, pp. 6901–6947

    work page 2010

Show all 73 references
  1. [8]

    Clary, J

    Xie, Z., P. Clary, J. Dao, P. Morais, J. Hurst, and M. Panne. Learning locomotion skills for cassie: Iterative design and sim- to-real. InConference on Robot Learning. PMLR, 2020, pp. 317–329

    2020

  2. [9]

    Li, Y ., J. Li, W. Fu, and Y . Wu. Learning agile bipedal motions on a quadrupedal robot. In2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 9735– 9742. Prepared usingTRR.cls 18 XX(X)

    2024

  3. [10]

    Bellegarda, G., Y . Chen, Z. Liu, and Q. Nguyen. Robust high- speed running for quadruped robots via deep reinforcement learning. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 10364– 10370

    2022

  4. [11]

    Peng, X. B., E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine. Learning agile robotic locomotion skills by imitating animals.arXiv preprint arXiv:2004.00784

    2004 arXiv

  5. [12]

    Allshire, V

    Handa, A., A. Allshire, V . Makoviychuk, A. Petrenko, R. Singh, J. Liu, D. Makoviichuk, K. Van Wyk, A. Zhurkevich, B. Sundaralingam, et al. Dextreme: Transfer of agile in- hand manipulation from simulation to reality. In2023 IEEE International Conference on Robotics and Automa...

    2023

  6. [13]

    Mehta, B., M. Diaz, F. Golemo, C. J. Pal, and L. Paull. Active domain randomization. InConference on Robot Learning. PMLR, 2020, pp. 1162–1176

    2020

  7. [14]

    Ramos, F., R. C. Possas, and D. Fox. Bayessim: adaptive domain randomization via probabilistic inference for robotics simulators.arXiv preprint arXiv:1906.01728

    1906 arXiv

  8. [15]

    Zhang, T

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

    arXiv

  9. [16]

    Arndt, and V

    Tiboni, G., K. Arndt, and V . Kyrki. DROPO: Sim-to-real transfer with offline domain randomization.Robotics and Autonomous Systems, V ol. 166, 2023, p. 104432

    2023

  10. [17]

    Ajay, A., J. Wu, N. Fazeli, M. Bauza, L. P. Kaelbling, J. B. Tenenbaum, and A. Rodriguez. Augmenting physical simulators with stochastic neural networks: Case study of planar pushing and bouncing. In2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS...

    2018

  11. [18]

    Zhang, D

    Jiang, Y ., T. Zhang, D. Ho, Y . Bai, C. K. Liu, S. Levine, and J. Tan. Simgan: Hybrid simulator identification for domain adaptation via adversarial reinforcement learning. In2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 2884–2890

    2021

  12. [19]

    Golemo, F., A. A. Taiga, A. Courville, and P.-Y . Oudeyer. Sim- to-real transfer with neural-augmented robot simulation. In Conference on Robot Learning. PMLR, 2018, pp. 817–828

    2018

  13. [20]

    Hwangbo, J., J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots.Science Robotics, V ol. 4, No. 26, 2019, p. eaau5872

    2019

  14. [21]

    Fey, N., G. B. Margolis, M. Peticco, and P. Agrawal. Bridging the Sim-to-Real Gap for Athletic Loco-Manipulation.arXiv preprint arXiv:2502.10894

    arXiv

  15. [22]

    Miller, A., F. Yu, M. Brauckmann, and F. Farshidian. High- Performance Reinforcement Learning on Spot: Optimizing Simulation Parameters with Distributional Measures.arXiv preprint arXiv:2504.17857

    arXiv

  16. [23]

    He, T., J. Gao, W. Xiao, Y . Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan, et al. Asap: Aligning simulation and real-world physics for learning agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143

    arXiv

  17. [24]

    Sobanbabu, N., G. He, T. He, Y . Yang, and G. Shi. Sampling- based system identification with active exploration for legged robot sim2real learning.arXiv preprint arXiv:2505.14266

    arXiv

  18. [25]

    Tsai, Y .-Y ., H. Xu, Z. Ding, C. Zhang, E. Johns, and B. Huang. Droid: Minimizing the reality gap using single-shot human demonstration.IEEE Robotics and Automation Letters, V ol. 6, No. 2, 2021, pp. 3168–3175

    2021

  19. [26]

    Sontakke, N., H. Chae, S. Lee, T. Huang, D. W. Hong, and S. Hal. Residual physics learning and system identification for sim-to-real transfer of policies on buoyancy assisted legged robots. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE...

    2023

  20. [27]

    Escontrela, D

    Wu, P., A. Escontrela, D. Hafner, P. Abbeel, and K. Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning. PMLR, 2023, pp. 2226–2240

    2023

  21. [28]

    Krause, and M

    Li, C., A. Krause, and M. Hutter. Offline Robotic World Model: Learning Robotic Policies without a Physics Simulator.arXiv preprint arXiv:2504.16680

  22. [29]

    Hanna, J. P., S. Desai, H. Karnan, G. Warnell, and P. Stone. Grounded action transformation for sim-to-real reinforcement learning.Machine Learning, V ol. 110, No. 9, 2021, pp. 2469– 2499

    2021

  23. [30]

    Smith, L., J. C. Kew, X. B. Peng, S. Ha, J. Tan, and S. Levine. Legged robots that keep on learning: Fine-tuning locomotion policies in the real world. In2022 international conference on robotics and automation (ICRA). IEEE, 2022, pp. 1593–1599

    2022

  24. [31]

    Song, X., Y . Yang, K. Choromanski, K. Caluwaerts, W. Gao, C. Finn, and J. Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 3769–3776

    2020

  25. [32]

    Hwangbo, L

    Lee, J., J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, V ol. 5, No. 47, 2020, p. eabc5986

    2020

  26. [33]

    Ramos, G

    Muratore, F., F. Ramos, G. Turk, W. Yu, M. Gienger, and J. Peters. Robot learning from randomized simulations: A review.Frontiers in Robotics and AI, V ol. 9, 2022, p. 799893

    2022

  27. [34]

    Ju, H., R. Juan, R. Gomez, K. Nakamura, and G. Li. Transferring policy of deep reinforcement learning from simulation to reality for robotics.Nature Machine Intelligence, V ol. 4, No. 12, 2022, pp. 1077–1087

    2022

  28. [35]

    Jenelten, S

    Grandia, R., F. Jenelten, S. Yang, F. Farshidian, and M. Hutter. Perceptive locomotion through nonlinear model-predictive control.IEEE Transactions on Robotics, V ol. 39, No. 5, 2023, pp. 3402–3421

    2023

  29. [36]

    Miki, T., J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild.Science Robotics, V ol. 7, No. 62, 2022, p. eabk2822. Prepared usingTRR.cls Bjelonic et al. 19

    2022

  30. [37]

    Xue, H., C. Pan, Z. Yi, G. Qu, and G. Shi. Full-order sampling- based mpc for torque-level locomotion control via diffusion- style annealing.arXiv preprint arXiv:2409.15610

    arXiv

  31. [38]

    Rudin, N., J. He, J. Aurand, and M. Hutter. Parkour in the Wild: Learning a General and Extensible Agile Locomotion Policy Using Multi-expert Distillation and RL Fine-tuning.arXiv preprint arXiv:2505.11164

    arXiv

  32. [39]

    Wolski, P

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

    arXiv

  33. [40]

    Ding, and X

    Liu, Y ., J. Ding, and X. Liu. Ipo: Interior-point policy optimization under constraints. InProceedings of the AAAI conference on artificial intelligence, V ol. 34. 2020, pp. 4940– 4947

    2020

  34. [41]

    Kim, Y ., H. Oh, J. Lee, J. Choi, G. Ji, M. Jung, D. Youm, and J. Hwangbo. Not only rewards but also constraints: Applications on legged robot locomotion.IEEE Transactions on Robotics, V ol. 40, 2024, pp. 2984–3003

    2024

  35. [42]

    Kaelbling, L. P., M. L. Littman, and A. R. Cassandra. Planning and acting in partially observable stochastic domains.Artificial intelligence, V ol. 101, No. 1-2, 1998, pp. 99–134

    1998

  36. [43]

    Hausknecht, M. J. and P. Stone. Deep Recurrent Q-Learning for Partially Observable MDPs. InAAAI fall symposia, V ol. 45. 2015, p. 141

    2015

  37. [44]

    Andrychowicz, P

    Pinto, L., M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel. Asymmetric actor critic for image-based robot learning.arXiv preprint arXiv:1710.06542

    arXiv

  38. [45]

    Hoeller, P

    Rudin, N., D. Hoeller, P. Reist, and M. Hutter. Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning. InProceedings of the 5th Conference on Robot Learning,Proceedings of Machine Learning Research, V ol. 164 (A. Faust, D. Hsu, and G. Neumann, eds.)....

    2022

  39. [46]

    Ji, G., J. Mun, H. Kim, and J. Hwangbo. Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion.IEEE Robotics and Automation Letters, V ol. 7, No. 2, 2022, pp. 4630–4637

    2022

  40. [47]

    Shin, Y .-H., T.-G. Song, G. Ji, and H.-W. Park. Actuator- constrained reinforcement learning for high-speed quadrupedal locomotion.arXiv preprint arXiv:2312.17507

    arXiv

  41. [48]

    Bellegarda, and A

    Shafiee, M., G. Bellegarda, and A. Ijspeert. Manyquadrupeds: Learning a single locomotion policy for diverse quadruped robots. In2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 3471–3477

    2024

  42. [49]

    Ma, Y . J., W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Jayaraman, Y . Zhu, L. Fan, and A. Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931

    arXiv

  43. [50]

    Gupta, J

    Eysenbach, B., A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills without a reward function.arXiv preprint arXiv:1802.06070

    arXiv

  44. [51]

    Zhang, E

    Yang, Y ., T. Zhang, E. Coumans, J. Tan, and B. Boots. Fast and efficient locomotion via learned gait transitions. InConference on robot learning. PMLR, 2022, pp. 773–783

    2022

  45. [52]

    Wensing, P. M., A. Wang, S. Seok, D. Otten, J. Lang, and S. Kim. Proprioceptive actuator design in the mit cheetah: Impact mitigation and high-bandwidth physical interaction for dynamic legged robots.Ieee transactions on robotics, V ol. 33, No. 3, 2017, pp. 509–522

    2017

  46. [53]

    Flayols, A

    Fadini, G., T. Flayols, A. Del Prete, N. Mansard, and P. Sou`eres. Computational design of energy-efficient legged robots: Optimizing for size and actuators. In2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9898–9904

    2021

  47. [54]

    Chane-Sane, L

    Roux, C., E. Chane-Sane, L. De Matte ¨ıs, T. Flayols, J. Manhes, O. Stasse, and P. Sou`eres. Constrained Reinforcement Learning for Unstable Point-Feet Bipedal Locomotion Applied to the Bolt Robot.arXiv preprint arXiv:2508.02194

    arXiv

  48. [55]

    Ragazzo, G

    Ferrari, S., P. Ragazzo, G. Dilevrano, and G. Pellegrino. Flux and loss map based evaluation of the efficiency map of synchronous machines.IEEE Transactions on Industry Applications, V ol. 59, No. 2, 2022, pp. 1500–1509

    2022

  49. [56]

    Vicari, F

    Valsecchi, G., A. Vicari, F. Tischhauser, M. Garabini, and M. Hutter. Accurate power consumption estimation method makes walking robots energy efficient and quiet. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 13282–13288

    2024

  50. [57]

    Nomura, M. and M. Shibata. cmaes : A Simple yet Practical Python Library for CMA-ES.arXiv preprint arXiv:2402.01373

  51. [58]

    Tahmassebi, W

    Telikani, A., A. Tahmassebi, W. Banzhaf, and A. H. Gandomi. Evolutionary machine learning: A survey.ACM Computing Surveys (CSUR), V ol. 54, No. 8, 2021, pp. 1–35

    2021

  52. [59]

    The CMA evolution strategy: A tutorial.arXiv preprint arXiv:1604.00772

    Hansen, N. The CMA evolution strategy: A tutorial.arXiv preprint arXiv:1604.00772

    arXiv

  53. [60]

    Hansen, N. and A. Ostermeier. Completely derandomized self- adaptation in evolution strategies.Evolutionary computation, V ol. 9, No. 2, 2001, pp. 159–195

    2001

  54. [61]

    Rudin, L

    Valsecchi, G., N. Rudin, L. Nachtigall, K. Mayer, F. Tis- chhauser, and M. Hutter. Barry: a high-payload and agile quadruped robot.IEEE Robotics and Automation Letters, V ol. 8, No. 11, 2023, pp. 6939–6946

    2023

  55. [62]

    Gehring, D

    Hutter, M., C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, et al. Anymal-a highly mobile and dynamic quadrupedal robot. InIEEE/RSJ International conference on intelligent robots and systems (IROS). 2016, pp. 38–44

    2016

  56. [63]

    Watanabe, R., T. Miki, F. Shi, Y . Kadokawa, F. Bjelonic, K. Kawaharazuka, A. Cramariuc, and M. Hutter. Learning quiet walking for a small home robot.arXiv preprint arXiv:2502.10983

    arXiv

  57. [64]

    Bellicoso, C. D., K. Kr ¨amer, M. St¨auble, D. Sako, F. Jenelten, M. Bjelonic, and M. Hutter. Alma-articulated locomotion and manipulation for a torque-controllable robot. In2019 International conference on robotics and automation (ICRA). IEEE, 2019, pp. 8477–8483. Prepared us...

    2019

  58. [65]

    Cramariuc, F

    Ma, Y ., A. Cramariuc, F. Farshidian, and M. Hutter. Learning coordinated badminton skills for legged manipulators.Science Robotics, V ol. 10, No. 102, 2025, p. eadu3922

    2025

  59. [66]

    Buehler, M

    Spiridonov, A., F. Buehler, M. Berclaz, V . Schelbert, J. Geurts, E. Krasnova, E. Steinke, J. Toma, J. Wuethrich, R. Polat, et al. Spacehopper: A small-scale legged robot for exploring low- gravity celestial bodies. In2024 IEEE International Conference on Robotics and Automati...

    2024

  60. [67]

    Hildebrandt, K

    Arnold, M., L. Hildebrandt, K. Janssen, E. Ongan, P. B ¨urge, ´A. G. G ´abriel, J. Kennedy, R. Lolla, Q. Oppliger, M. Schaaf, et al. LEV A: A high-mobility logistic vehicle with legged suspension.arXiv preprint arXiv:2503.10028

    arXiv

  61. [68]

    Eugster, N

    Leuthard, S., T. Eugster, N. Faesch, R. Feingold, C. Flynn, M. Fritsche, N. H ¨urlimann, E. Morbach, F. Tischhauser, M. M ¨uller, et al. Magnecko: Design and Control of a Quadrupedal Magnetic Climbing Robot. InClimbing and Walking Robots Conference. Springer, 2024, pp. 55–67

    2024

  62. [69]

    Zhang, F

    He, J., C. Zhang, F. Jenelten, R. Grandia, M. B ¨Acher, and M. Hutter. Attention-Based Map Encoding for Learning Generalized Legged Locomotion.arXiv preprint arXiv:2506.09588

    arXiv

  63. [70]

    Gehring, and D

    Hottiger, G., C. Gehring, and D. Bellicoso. signal logger: Signal logging and plotting tools for robotics.https: //github.com/ANYbotics/signal_logger, 2020. Accessed: 2025-07-09

    2020

  64. [72]

    Bryce, C. M. and T. M. Williams. Comparative locomotor costs of domestic dogs reveal energetic economy of wolf-like breeds. Journal of Experimental Biology, V ol. 220, No. 2, 2017, pp. 312–321

    2017

  65. [73]

    Springob, L. and J. Holtz. High-bandwidth current control for torque-ripple compensation in PM synchronous machines. IEEE Transactions on industrial electronics, V ol. 45, No. 5, 2002, pp. 713–721. A Voltage–Limited Torque Bandwidth of a PMSM Even with perfect control, a PMSM’...

    2002

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.