Pith. sign in

REVIEW 4 major objections 5 minor 43 references

McARL:Morphology-Control-Aware Reinforcement Learning for Generalizable Quadrupedal Locomotion

T0 review · 4 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 claims that a single PPO policy trained on a Unitree Go1 in simulation, conditioned on a randomized 14-dimensional morphology/control vector and a history-aware curriculum, transfers zero-shot to a Go2 at 3.5 m/s, to a Mini…

desk verdict Useful ablation study with an overstated headline: the Go1-to-Go2 'zero-shot' result is likely interpolation inside the morphology randomization ranges, but the paper is honest enough about its limits to deserve a careful referee. read the letter →

arxiv 2505.18418 v1 pith:QWISJF3V submitted 2025-05-23 cs.RO

Prakhar Mishra , Amir Hossain Raj , Xuesu Xiao , Dinesh Manocha This is my paper
classification cs.RO
keywords quadrupedallocomotionmorphology-conditionedpolicyzero-shottransfermorphologyrandomizationcurriculumlearningPPOreinforcement
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

McARL is a reinforcement-learning method whose thesis is that a quadruped's morphology should be an input to the controller rather than environmental noise. The paper trains one PPO policy on the Unitree Go1 in simulation, feeding a randomly sampled 14-dimensional vector of link lengths, masses, joint ranges, and actuator properties into both the actor and critic through a learned embedding. The same policy is then tested directly on other quadrupeds, with reported zero-shot sim-to-sim transfer as high as 3.5 m/s on the Go2, 1.5 m/s on the Mini Cheetah, and stable standing on the A1, while the training robot itself reaches 6.0 m/s. Relative to PPO variants without morphology conditioning, McARL claims 44-150% higher transfer performance. If the claim is right, users could train once on an available quadruped and deploy the controller on a whole family of similar robots without per-robot retraining or hyperparameter tuning.

What carries the argument

The argument is carried by the morphology-conditioned actor-critic: a PPO policy whose inputs $x_t=[o_t,\kappa_t,z_m]$ include the 64-dimensional morphology embedding $z_m$ alongside observations and an RNN history term, so that both $\pi_\theta(a_t|o_t,h_t,m)$ and $V_\phi(s_t,m)$ are conditioned on the randomized body description at every forward pass. The second mechanism is the History-Aware Curriculum, which splits the command-velocity space into 4000 bins, uses an LSTM hidden state $H_t$ to predict $\hat r_{\mathrm{lin}}(b_i)$ and $\hat r_{\mathrm{ang}}(b_i)$, and updates bin weights via $w_{t+1}(b)=w_t(b)+\alpha(\hat r_{\mathrm{lin}}(b)+\hat r_{\mathrm{ang}}(b))$, steering learning toward commands the current morphology handles well. The morphology-distance analysis uses the weighted Euclidean distance $D_{\mathrm{weighted}}(i,j)$ of Eq. (15) as a diagnostic, and the paper's claim is that this distance predicts the drop in transfer performance.

What would settle it

Take the trained Go1 policy and run it zero-shot on a simulated Go2 whose base mass, link lengths, and torque limits are moved outside the Table 1 ranges; if it still tracks 3.5 m/s, morphology conditioning does more than interpolate, and if it collapses, the randomization envelope is the load-bearing assumption. A direct disconfirmation would also come from a real Go2 or Mini Cheetah run with no fine-tuning failing to walk, since the headline claim is that retraining is unnecessary.

Watch

Extended reading notes

Core claim

On the paper's own framing, the central discovery is that morphology is a learnable conditioning variable rather than randomized noise. The policy is defined over a parameterized MDP $M(m)=\langle S,A,P_m,r_m,\gamma\rangle$ with $m\sim\mathrm{Uniform}[m_{\min},m_{\max}]\subset\mathbb{R}^{14}$, and a two-layer encoder $z_m=f_\psi(m)=\mathrm{ELU}(W_2\mathrm{ELU}(W_1m+b_1)+b_2)\in\mathbb{R}^{64}$ turns each sampled morphology-control vector into a latent that is concatenated with the observation and history estimate in both the Gaussian actor and the value function. With this input, the usual clipped PPO objective produces a single policy that reaches $6.0\,\mathrm{m/s}$ on the Go1 in simulation and transfers zero-shot to $3.5\,\mathrm{m/s}$ on the Go2, $1.5\,\mathrm{m/s}$ on the Mini Cheetah, and a stable stance on the A1. The paper also claims that the history-aware curriculum, an RNN which predicts per-bin linear and angular rewards and reweights 4000 command-velocity bins, is important to this result, and it quantifies transfer loss with a normalized weighted 14-dimensional morphology distance, showing performance degrades as the trained and target robots grow apart in that space. On hardware, the method is demonstrated on the Go1 with some fine-tuning at 1.0-1.5 m/s command velocities.

Load-bearing premise

The zero-shot claim rests on every target robot's true parameters lying inside the hand-defined 14D randomization ranges of Table 1, so the transfer is interpolation inside a designer-chosen morphology envelope rather than extrapolation to unseen body plans.

Editorial extensions

If this is right

  • One controller trained on the Go1 can be dropped onto the Go2, Mini Cheetah, and A1 in simulation with no retraining, with the reported speeds of 3.5, 1.5, and stable standing respectively.
  • Morphology conditioning pays: relative to state-only or non-morphology PPO variants, McARL reports 44-150% higher transfer performance on Go2, Mini Cheetah, and A1.
  • Transfer quality is not arbitrary; it tracks the normalized weighted 14D morphology distance, so the method gives a quantitative predictor of when a new platform will work.
  • Adding history-awareness to the curriculum is part of the recipe; the ablations show morphology input alone transfers worse than morphology plus the history-aware curriculum.
  • The same training recipe can start from a different robot (Go2, Mini Cheetah, A1) and transfer to the remaining three, although the A1 transfer is the least stable.

Reading between the lines

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

  • The 14D randomization ranges define an interpolation envelope; a deployment robot whose parameters fall outside Table 1 would be extrapolation, and the paper's distance-vs-loss result suggests performance there is not guaranteed.
  • The morphology latent and the distance metric together suggest a practical adaptation loop: estimate the target's 14D parameters online or offline, feed them to the same policy, and use the weighted distance to predict whether fine-tuning is needed.
  • Because the method treats the body plan as a conditioning input, the same architecture should in principle extend to other legged families or manipulators, provided the 14D vector is replaced with an appropriate parameterization; the paper lists this as future work but does not demonstrate it.
  • The comparison between the morphology-only and morphology-plus-control-conditioned variants hints that including control gains in the latent can make the controller more conservative; a natural test would be separating morphology and control parameters to see which one actually carries transfer.
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

4 major / 5 minor

Summary. The paper proposes McARL, a PPO-based locomotion controller that augments the actor and critic inputs with a 64-dimensional morphology embedding derived from a randomized 14-dimensional morphology/control vector, and combines this with a history-aware curriculum learning scheme (HACL) that uses an LSTM to predict per-bin rewards and adapt command sampling. The experiments train on a Unitree Go1 simulation and report zero-shot sim-to-sim transfer to Go2 at 3.5 m/s, Mini Cheetah at about 1.5 m/s, and A1 at about 0.5 m/s, with 6.0 m/s on Go1; six PPO variants are ablated in Table 5. The paper also introduces a weighted morphology distance metric (Eq. 15) to relate transfer loss to morphology distance, and describes a real-world Go1 deployment after fine-tuning.

Significance. If the central claim were established—one policy trained on a single robot and transferred zero-shot to a genuinely unseen morphology—McARL would be a practically valuable step for quadruped locomotion and a useful extension of morphology-conditioned RL. The paper's strengths include a detailed ablation table, openly documented hyperparameters, a project page with videos, and an explicit distance-based analysis of transfer loss. However, as written, the headline result is not yet supported: the target robots appear to fall inside the hand-defined randomization envelope of Table 1, and the paper does not clarify whether the randomization changes the simulator dynamics or only the network input. The significance of the contribution therefore hinges on clarification and on out-of-distribution evaluation, which the current manuscript does not provide.

major comments (4)
  1. [Abstract, Section 3.1, Table 1, Eq. (15)] The zero-shot transfer claim is not supported as stated because the evaluation appears to be in-distribution interpolation, not extrapolation to unseen morphologies. Section 3.1 states that each of the 4000 training environments is assigned a fresh 14D vector m sampled uniformly from the Table 1 ranges, and Section 4.1 repeats that a new set of vectors is generated each episode. Table 1's ranges (e.g., base mass [3.3, 6.921] kg, thigh-to-calf length [0.2, 0.213] m, torque limits) are hand-defined from the same robot family—Go1, Go2, A1, Mini Cheetah—that is later used for evaluation. The paper never reports the nominal 14D vectors of Go2, A1, and Mini Cheetah or checks whether they lie inside the training support. If they do, the policy has been trained on a distribution covering these morphologies, so the 'trained on Go1' claim is misleading and the 3.5 m/s Go2 result is interpolation, not transfer to an unseen body plan. The paper's own Eq. (15) and Figure 4, which show transfer loss growing with weighted 14D distance, are consistent with this reading. The authors should either restrict the training randomization so that target robots are genuinely out of distribution, or reframe the claim as interpolation over a designer-specified morphology envelope.
  2. [Sections 3.1 and 4.1] It is never made explicit whether the sampled 14D morphology/control vector is used only as a conditioning input to the encoder or also to set the physical parameters of the simulated robot (link lengths, masses, torque limits, actuator gains). Section 3.1 and Eq. (2) describe only the generation of m and its encoding into z_m; Section 4.1 says 'a new set of morphology vectors are generated' but does not state that the Isaac Gym URDF models or actuator properties are modified accordingly. If the dynamics are not randomized, then all training is on the nominal Go1 dynamics and the Table 1 ranges are never exercised, so the 'morphology-aware' mechanism and the ablations in Table 5 do not test morphology generalization. If the dynamics are randomized, the authors must describe exactly how each of the 14 entries is applied to the simulation and confirm that the target URDFs used at evaluation correspond to points inside the training distribution. This ambiguity is load-bearing for every transfer result in the paper.
  3. [Table 5, Section 5.2, Figure 5] The identity of the best variant and the attribution of the result to the morphology embedding are unclear. Table 5 reports that P4 (state + morph + ctrl, HACL) achieves 6.0 m/s on Go1 and 3.47 m/s on Go2, slightly higher than P3 (5.98 and 3.43), while P3 is better on Mini Cheetah (1.47 vs 0.5) and A1 (0.5 vs 0.35). Section 5.2 and Figure 5 nonetheless declare P3 'overall best', but no aggregate metric over the four robots or statistical selection criterion is defined. Moreover, P1 (state + one-hot ID vector, no morphology, no HACL) reaches 3.42 m/s on Go2, nearly matching P3's 3.43 m/s, so the specific benefit of the learned morphology embedding over a simple robot-identity input is not established by this comparison. The authors should define the selection metric, report multiple seeds, and discuss the P1 result explicitly.
  4. [Section 5.4, Section 7] The real-world claim is scoped inconsistently. The introduction says the real-world demonstration required 'minimal finetuning', and Section 5.4 states that the authors 'fine-tuned the model to improve the robustness and stability' after observing unstable gait and forward falling. Section 7 correctly lists sim-to-real challenges and admits that fine-tuning is needed. This is acceptable for a sim-to-sim zero-shot claim, but the abstract's phrase 'without retraining or fine-tuning' should not be read as applying to the real-world deployment; the paper should state plainly that real-world deployment required additional tuning and that the zero-shot claim is limited to simulation.
minor comments (5)
  1. [Table 2] The reward expressions for r_lin and r_ang are identical in the table; the angular-velocity reward should presumably compare angular velocity to the commanded yaw rate. Please correct the typo.
  2. [Eq. (15), Section 5.3] The distance metric uses z_k(i) with k = 1,...,14, but z_m was defined in Eq. (2) as a 64-dimensional embedding. Clarify whether z in Eq. (15) is the raw 14D vector or the embedding; if it is the raw vector, avoid reusing z.
  3. [Eqs. (11)-(14), Section 3.4] The notation x_t is overloaded: Eq. (5) uses x_t for the policy input, while Eq. (11) redefines x_t as a one-hot bin indicator in R^4000. Use a different symbol, e.g., b_t, for the curriculum input.
  4. [References] Reference [39] is a self-citation with placeholder 'arXiv:2505.XXXX'; it should be completed or removed before publication.
  5. [Table numbering and cross-references] Section 5.1 refers to 'the results are elaborated in the table 3' when describing Table 5, and Section 5.2 similarly refers to 'table 3' for the ablation results. Renumber or fix the cross-references so that Table 5 is cited where the PPO variants are discussed.

Circularity Check

1 steps flagged · score 4.0 of 10

The Go2 zero-shot result is in-distribution interpolation because the 14D morphology randomization envelope is constructed from the very robots later used as transfer targets.

  1. fitted input called prediction [Abstract; Section 3.1 (Morphology Randomization and Encoder), Table 1; Section 5.3, Eq. (15)]
    "a single policy trained on a Unitree Go1 robot using McARL can be transferred to a different morphology (e.g., Unitree Go2 robot) and can achieve zero-shot transfer velocity of as high as 3.5 m/s without retraining or fine-tuning. ... Our goal ... select the right family of legged robots and based on those robots, we should define the morphology and control parameters ... In our case we consider these robots: Unitree Go1, Go2, Mini-cheetah and A1. ... Based on Table 1 parameters ranges, we generate a separate random vector of dimension 14 for all 4000 training environments ..."

    The randomization support is explicitly built from the same robot family used as transfer targets, so Go2's nominal parameters are inside the training distribution by construction. Every episode samples m from this support and conditions actor and critic on it (Eqs. 5-7). The abstract calls Go2 a 'different morphology' and the result 'zero-shot,' but the policy has been trained on inputs covering Go2-like parameter vectors; the evaluation is interpolation inside a designer-specified envelope, not prediction for an unseen body plan. The paper's own Section 5.3, Eq. (15), and Section 7 state that transfer loss grows with weighted 14D morphology distance, confirming coverage-dependent behavior rather than extrapolation.

full rationale

The PPO training objective (Eqs. 8-10), the morphology encoder (Eq. 2), and the policy architecture are not self-referential; they are standard learning components. The main circularity is in the experiment framing: the morphology range is selected from the target robots, and then transfer to those robots is advertised as zero-shot. That is a fitted input called a prediction, though the achieved speeds still require real learning, and the baseline comparisons give the paper independent content. The self-citation [39] for HACL is minor and not load-bearing; Section 3.4 specifies the HACL equations. A separate internal inconsistency appears in Section 4.2 ('we have also trained 400 million timestep in simulation using 4000 environments of A1, Go1, Go2, and MIT mini cheetah robots'), which, if read literally, contradicts the 'trained on Go1 only' premise; this is a support/correctness issue rather than a circularity. Overall score 4.0: one central claim is partially reduced by construction, but most of the pipeline is independently evaluated.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-made constructs: the 14D randomization ranges in Table 1, chosen to bracket the four evaluation robots, and the reward and curriculum hyperparameters. No new physical entities are introduced; the morphology latent z_m is an internal network representation. The paper is an empirical RL study rather than a derivation, so the ledger mainly tracks the manually specified ranges and weights.

free parameters (3)
  • Morphology and control randomization ranges = Table 1 bounds, e.g., base mass [3.3, 6.921] kg, joint stiffness [20, 30] Nm/rad
    Hand-selected to bracket Go1, Go2, A1 and Mini Cheetah; all zero-shot evaluation robots fall inside these bounds, making the transfer test an interpolation check.
  • Reward weights = e.g., rlin=3.0, rang=3.0, raction rate=0.2, renergy efficiency=1e-7
    Chosen by hand and inherited from open-source bases [1] and [3]; no sensitivity analysis is provided, and the reward directly shapes the gait and transfer behavior.
  • HACL bin size and weight update parameters = bin size 0.5, design space (20,10,20), weight update 0.2*rlin + rang/2, RNN hidden size 128
    Hand-set curriculum parameters; the curriculum is a central component of the best-performing variant, and no ablation varies these values.
assumptions (4)
  • domain assumption IsaacGym with the provided URDFs faithfully models Go1, Go2, A1 and Mini Cheetah dynamics, and the sim results are representative of hardware behavior.
    Invoked in Section 4.2; all zero-shot transfer numbers are sim-to-sim, and Section 5.4 admits real deployment needs fine-tuning.
  • ad hoc to paper The 14D morphology/control vector is a sufficient descriptor for cross-morphology transfer.
    Invoked in Section 3.1 and Table 1; unmodeled factors such as foot geometry, friction, center-of-mass offsets and actuator latency are not included.
  • domain assumption The evaluation robots lie within the sampled morphology randomization distribution.
    Table 1 ranges are derived from the four robots; Section 5.3 shows transfer loss grows with weighted 14D distance, so success depends on staying inside the envelope.
  • standard math PPO with the clipped surrogate loss converges to a near-optimal policy for the parameterized MDP family.
    Section 3.3 states PPO is not modified except for conditioning on x_t; the method inherits PPO's convergence assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of McARL:Morphology-Control-Aware Reinforcement Learning for Generalizable Quadrupedal Locomotion." pith.science (2026). https://pith.science/paper/QWISJF3V

@misc{pith2026250518418,
  author       = {Pith},
  title        = {Pith review of: McARL:Morphology-Control-Aware Reinforcement Learning for Generalizable Quadrupedal Locomotion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QWISJF3V}},
  note         = {Machine review of arXiv:2505.18418}
}
read the original abstract

We present Morphology-Control-Aware Reinforcement Learning (McARL), a new approach to overcome challenges of hyperparameter tuning and transfer loss, enabling generalizable locomotion across robot morphologies. We use a morphology-conditioned policy by incorporating a randomized morphology vector, sampled from a defined morphology range, into both the actor and critic networks. This allows the policy to learn parameters that generalize to robots with similar characteristics. We demonstrate that a single policy trained on a Unitree Go1 robot using McARL can be transferred to a different morphology (e.g., Unitree Go2 robot) and can achieve zero-shot transfer velocity of up to 3.5 m/s without retraining or fine-tuning. Moreover, it achieves 6.0 m/s on the training Go1 robot and generalizes to other morphologies like A1 and Mini Cheetah. We also analyze the impact of morphology distance on transfer performance and highlight McARL's advantages over prior approaches. McARL achieves 44-150% higher transfer performance on Go2, Mini Cheetah, and A1 compared to PPO variants.

Figures

Figures reproduced from arXiv: 2505.18418 by the authors.

Figure 1
Figure 1. Overview of our Morphology-Control-Aware Reinforcement Learning (McARL), a novel [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Morphology embedded policy controller testing in simulator on go1 robot (row 1) and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The first image of each row represents on which ronot McARL has been trained and the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Parallelized training for go1, go2, mini cheetah and A1 robot (total 4000 environments) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: PPO combination of P3 PPO + Morphology (without control params) + HACL performs [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 22 canonical work pages

  1. [1]

    G. B. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal. Rapid locomotion via rein- forcement learning.The International Journal of Robotics Research, 43(4):572–587, 2024

    work page 2024

  2. [2]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034, 2021

    arXiv 2021

  3. [3]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. InConference on Robot Learning, pages 91–100. PMLR, 2022

    2022

  4. [4]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

    2020

  5. [5]

    Bledt, M

    G. Bledt, M. J. Powell, B. Katz, J. Di Carlo, P. M. Wensing, and S. Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018

    work page 2018

  6. [6]

    Bosworth, J

    W. Bosworth, J. Whitney, S. Kim, and N. Hogan. Robot locomotion on hard and soft ground: Measuring stability and ground properties in-situ. In2016 IEEE international conference on Robotics and automation (ICRA), pages 3582–3589. IEEE, 2016

    work page 2016

  7. [7]

    Y . Ding, A. Pandala, and H.-W. Park. Real-time model predictive control for versatile dynamic motions in quadrupedal robots. In2019 International Conference on Robotics and Automation (ICRA), pages 8484–8490. IEEE, 2019

    work page 2019

  8. [8]

    Herzog, S

    A. Herzog, S. Schaal, and L. Righetti. Structured contact force optimization for kino-dynamic motion generation. In2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2703–2710. IEEE, 2016

    work page 2016

Show all 43 references
  1. [9]

    W. Yu, V . C. Kumar, G. Turk, and C. K. Liu. Sim-to-real transfer for biped locomotion. In 2019 ieee/rsj international conference on intelligent robots and systems (iros), pages 3503–

    2019

  2. [10]

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, and V . Vanhoucke. Sim- to-real: Learning agile locomotion for quadruped robots.arXiv preprint arXiv:1804.10332, 2018

    2018 arXiv

  3. [11]

    G. Feng, H. Zhang, Z. Li, X. B. Peng, B. Basireddy, L. Yue, Z. Song, L. Yang, Y . Liu, K. Sreenath, et al. Genloco: Generalized locomotion controllers for quadrupedal robots. In Conference on Robot Learning, pages 1893–1903. PMLR, 2023

    1903

  4. [12]

    Sferrazza, D.-M

    C. Sferrazza, D.-M. Huang, F. Liu, J. Lee, and P. Abbeel. Body transformer: Leveraging robot embodiment for policy learning.arXiv preprint arXiv:2408.06316, 2024

    2024 arXiv

  5. [13]

    Shafiee, G

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

    2024

  6. [14]

    Bohlinger, G

    N. Bohlinger, G. Czechmanowski, M. Krupka, P. Kicki, K. Walas, J. Peters, and D. Tateo. One policy to run them all: an end-to-end learning approach to multi-embodiment locomotion. arXiv preprint arXiv:2409.06366, 2024

    2024

  7. [15]

    Fahmi, M

    S. Fahmi, M. Focchi, A. Radulescu, G. Fink, V . Barasuol, and C. Semini. Stance: Locomotion adaptation over soft terrain.IEEE Transactions on Robotics, 36(2):443–457, 2020

    2020

  8. [16]

    Kuindersma, R

    S. Kuindersma, R. Deits, M. Fallon, A. Valenzuela, H. Dai, F. Permenter, T. Koolen, P. Marion, and R. Tedrake. Optimization-based locomotion planning, estimation, and control design for the atlas humanoid robot.Autonomous robots, 40:429–455, 2016

    2016

  9. [17]

    H. Dai, A. Valenzuela, and R. Tedrake. Whole-body motion planning with centroidal dynamics and full kinematics. In2014 IEEE-RAS International Conference on Humanoid Robots, pages 295–302. IEEE, 2014. 11

    2014

  10. [18]

    D. Kim, J. Di Carlo, B. Katz, G. Bledt, and S. Kim. Highly dynamic quadruped locomotion via whole-body impulse control and model predictive control.arXiv preprint arXiv:1909.06586, 2019

    1909 arXiv

  11. [19]

    Tobin, R

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30. IEEE, 2017

    2017

  12. [20]

    Z. Xie, X. Da, M. Van de Panne, B. Babich, and A. Garg. Dynamics randomization revisited: A case study for quadrupedal locomotion. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4955–4961. IEEE, 2021

    2021

  13. [21]

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In2018 IEEE international conference on robotics and automation (ICRA), pages 3803–3810. IEEE, 2018

    2018

  14. [22]

    Nachum, M

    O. Nachum, M. Ahn, H. Ponte, S. Gu, and V . Kumar. Multi-agent manipulation via locomotion using hierarchical sim2real.arXiv preprint arXiv:1908.05224, 2019

    1908 arXiv

  15. [23]

    Luo and K

    J. Luo and K. Hauser. Robust trajectory optimization under frictional contact with iterative learning.Autonomous Robots, 41:1447–1461, 2017

    2017

  16. [24]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learn- ing agile and dynamic motor skills for legged robots.Science Robotics, 4(26):eaau5872, 2019

    2019

  17. [25]

    Y . Zhao, T. Wu, Y . Zhu, X. Lu, J. Wang, H. Bou-Ammar, X. Zhang, and P. Du. Zsl-rppo: Zero-shot learning for quadrupedal locomotion in challenging terrains using recurrent proximal policy optimization.arXiv preprint arXiv:2403.01928, 2024

    2024 arXiv

  18. [26]

    H. Li, W. Yu, T. Zhang, and P. M. Wensing. Zero-shot retargeting of learned quadruped lo- comotion policies using hybrid kinodynamic model predictive control. In2022 IEEE/RSJ In- ternational Conference on Intelligent Robots and Systems (IROS), pages 11971–11977. IEEE, 2022

    2022

  19. [27]

    X. Pan, T. Zhang, B. Ichter, A. Faust, J. Tan, and S. Ha. Zero-shot imitation learning from demonstrations for legged robot visual navigation. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 679–685. IEEE, 2020

    2020

  20. [28]

    Rafiei, L

    M. Rafiei, L. Gauthier, H. Adeli, and D. Takabi. A survey of deep learning techniques for malware analysis.IEEE Transactions on Neural Networks and Learning Systems, 2022

    2022

  21. [29]

    Scarselli, M

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini. The graph neural network model.IEEE transactions on neural networks, 20(1):61–80, 2008

    2008

  22. [30]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  23. [31]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. De- hghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transform- ers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  24. [32]

    Huang, I

    W. Huang, I. Mordatch, and D. Pathak. One policy to control them all: Shared modular policies for agent-agnostic control. InInternational Conference on Machine Learning, pages 4455–4464. PMLR, 2020

    2020

  25. [33]

    A. S. Chiappa, A. Marin Vargas, and A. Mathis. Dmap: a distributed morphological atten- tion policy for learning to locomote with a changing body.Advances in Neural Information Processing Systems, 35:37214–37227, 2022

    2022

  26. [34]

    Gupta, L

    A. Gupta, L. Fan, S. Ganguli, and L. Fei-Fei. Metamorph: Learning universal controllers with transformers.arXiv preprint arXiv:2203.11931, 2022

    2022 arXiv

  27. [35]

    Bengio, J

    Y . Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. InProceedings of the 26th annual international conference on machine learning, pages 41–48, 2009. 12

    2009

  28. [36]

    X. Wang, Y . Chen, and W. Zhu. A survey on curriculum learning.IEEE transactions on pattern analysis and machine intelligence, 44(9):4555–4576, 2021

    2021

  29. [37]

    Matiisen, A

    T. Matiisen, A. Oliver, T. Cohen, and J. Schulman. Teacher–student curriculum learning.IEEE transactions on neural networks and learning systems, 31(9):3732–3740, 2019

    2019

  30. [38]

    Aractingi, P.-A

    M. Aractingi, 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, 13(1):11945, 2023

    2023

  31. [39]

    Mishra, A

    P. Mishra, A. H. Raj, X. Xiao, and D. Manocha. Hacl: History-aware curriculum learning for fast locomotion.arXiv preprint arXiv:2505.XXXX, 2025. Preprint available

    2025

  32. [40]

    Rudin, D

    N. Rudin, D. Hoeller, M. Bjelonic, and M. Hutter. Advanced skills by learning locomotion and local navigation end-to-end. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2497–2503. IEEE, 2022

    2022

  33. [41]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    2021 arXiv

  34. [42]

    Zhao and Y

    Y . Zhao and Y . Gu. A non-periodic planning and control framework of dynamic legged loco- motion.International Journal of Intelligent Robotics and Applications, 4(1):95–108, 2020

    2020

  35. [43]

    X. Chen, J. Hu, C. Jin, L. Li, and L. Wang. Understanding domain randomization for sim-to- real transfer.arXiv preprint arXiv:2110.03239, 2021. 13

    2021 arXiv

Pith tools

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