Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Learning to Dock: A Simulation-based Study on Closing the Sim2Real Gap in Autonomous Underwater Docking

T0 review · 4 major / 6 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 For AUV docking, a naive RL policy survives payload changes in simulation, while domain randomization and history add only marginal gains under extreme loads.

desk verdict A clean sim-only comparison of DR and history for AUV docking under payload shifts, with a plausible negative result that needs more statistical rigor and a softened real-world claim. read the letter →

arxiv 2506.17823 v1 pith:3Z4VADGD submitted 2025-06-21 cs.RO cs.AIcs.LG

Kevin Chang , Rakesh Vivekanandan , Noah Pragin , Sean Bullock , Geoffrey Hollinger This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords autonomousunderwaterdockingsim2realgapreinforcementlearningdomainrandomizationpayloadrobustnesshistory-conditionedpoliciesrobotics
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

This paper asks whether the standard recipe for closing the sim-to-real gap—domain randomization and memory-based policies—is actually needed to dock an autonomous underwater vehicle (AUV) when the vehicle's payload changes. In a high-throughput simulator, the authors train a naive PPO policy, two domain-randomized variants, and one that conditions on recent state history, then evaluate all four under easy, medium, and hard payload disturbances. The central result is that the naive policy remains effective across the realistic payload range; domain randomization only shows marginal gains under the hardest payload, and adding history gives a small further edge but higher variance. The authors conclude that, given pose and twist estimates, a docking policy does not require randomization or memory to absorb mass-distribution shifts, and suggest real-world deployment start with the simplest policy. A sympathetic reader would care because this recommends a cheaper, simpler path to robust docking controllers.

What carries the argument

The comparison rests on four policy configurations sharing one MLP architecture and differing only in input dimension: a naive policy with one-step observations; two domain-randomized policies whose training randomly samples payload mass (up to 2.5 kg at 0.1 m radius, or 5.0 kg at 0.3 m radius); and the larger-randomization policy fed a concatenated history of three observations. The reward is a weighted sum of exponential penalties on position and orientation error. The load-bearing comparison is the positional and angular error curves across easy, medium, and hard payload evaluation scenarios.

What would settle it

Run the same naive policy on a real AUV carrying a 7 kg payload placed 0.3 m along its x-axis and compare positional error over time against the simulated hard-payload curve; agreement would confirm the claim, and large divergence would show the simulated payload model missed the real sim2real gap.

Watch

Extended reading notes

Core claim

The paper's claim is that for 6-DOF AUV docking, a policy trained with plain PPO on a fixed-payload simulator generalizes zero-shot to out-of-distribution payloads nearly as well as policies deliberately trained for robustness. Across all three evaluation scenarios, positioning error stays roughly equal for all configurations; under a 7 kg payload placed 0.3 m along the x-axis, the large-domain-randomization policy edges out the naive policy, and adding a history of three observations to that policy yields a marginal further improvement while also increasing variance. The authors interpret this as evidence that shifting a vehicle's mass distribution does not change the optimal mapping from state to thruster commands enough to require explicit robustness training, and therefore that DR and memory should be treated as conservative add-ons for extreme conditions rather than prerequisites. They explicitly limit the study to simulation, so the claim is about simulated robustness under modeled payload disturbances.

Load-bearing premise

The study's recommendation rests on the premise that the simulated payload disturbance—a sampled mass placed at an offset—captures the real-world sim2real gap caused by attaching payloads to an AUV; the paper was evaluated entirely in simulation, so this fidelity is untested.

Editorial extensions

If this is right

  • A practitioner can begin real-world docking trials with a naively trained PPO policy and reserve domain randomization for observed failures under extreme payloads.
  • Training budgets on similar underwater docking tasks can be reduced by skipping DR and history augmentation for nominal missions.
  • The finding that history adds variance suggests memory-based architectures should only be introduced when DR is already in place and extreme loads are expected.
  • The evaluation protocol—normal, medium, and hard payload shifts—offers a cheap, reproducible robustness test for docking controllers before costly sea trials.
  • If the explanation holds, payload-induced sim2real gaps are dominated by dynamics shifts that feedback can correct, not by state-estimation or actuation mismatch.

Reading between the lines

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

  • The same reasoning would predict that a naive policy also tolerates mild unmodeled drag or thruster degradation, since those also appear as state-dependent dynamics changes; testing that would extend the claim beyond payloads.
  • Because the paper's DR samples a point mass offset, it does not generate added mass or asymmetric hydrodynamic drag; a real payload's wetted geometry could create effects the sim's disturbance model misses, so the strongest version of the conclusion should be read as about mass-shift disturbances only.
  • A useful benchmark extension would be to evaluate the same four policies under payloads that also change the vehicle's inertia tensor and center of buoyancy, not just center of mass, to see when the naive policy starts to fail.
  • The angular-error result, where policies misalign along the Z-axis, suggests that tuning the orientation reward weight or randomizing rotational offsets during training might close the remaining gap more directly than adding history.
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

4 major / 6 minor

Summary. The paper presents a simulation-based study of reinforcement learning docking controllers for a BlueROV2 Heavy AUV. Four PPO policy configurations (naive, small domain randomization, large DR, large DR with state history) are trained in NVIDIA Isaac Sim and evaluated in simulation under three payload scenarios (easy, medium, hard). The central finding is that a naively trained policy is largely as effective as DR/history-based policies for the simulated docking task, with DR and history providing only marginal improvements under the hardest payload, and the paper concludes by recommending that real-world deployments start with the naive policy. The manuscript explicitly states that no real-world experiments were conducted.

Significance. If taken at face value, the controlled four-way comparison provides a useful data point on whether domain randomization and history conditioning are needed for AUV docking under mass-distribution shifts, and the finding that a naive policy is robust to such shifts is non-obvious. The study's strengths are the clear experimental design, multiple training seeds, and a well-specified payload-shift disturbance. However, the evidence is thin: 20 episodes per condition with no error bars or statistical tests, a single disturbance channel, and no validation against real vehicle data or a higher-fidelity model. The contribution is therefore best framed as a preliminary simulation study rather than a demonstration of how to close the sim2real gap, and the real-world recommendation in Section VI goes beyond what the evidence supports.

major comments (4)
  1. [Section III.C and Section VI] The paper acknowledges that it is simulation-only, yet the concluding recommendation is a real-world deployment strategy: 'when testing a learned docking controller in the real world, it may be most reasonable to start with a naively trained policy.' This recommendation is not supported by the evidence, because the dynamics model (Section III.A.2) relies on MuJoCo's simple inertial box model and zero-order thruster dynamics, and the only disturbance studied is a static point-mass payload. No validation against real BlueROV2 data or a higher-fidelity hydrodynamic model is presented, so the phrase 'realistic payloads' is an unsupported premise for the paper's central claim. A concrete remedy is to add a validation experiment against real-vehicle data or, at minimum, to restrict all conclusions to statements about the simulator and remove the real-world recommendation.
  2. [Section IV.B and Figures 4-9] Each evaluation condition uses only 20 episodes (Section IV.B), yet Figures 4-9 show no error bars, confidence bands, or per-episode distributions, and the text describes performance differences qualitatively ('roughly equally well', 'marginally improve', 'consistently perform the worst'). No significance tests or effect sizes are reported. Because the central claim is that DR and history provide only marginal benefits, the absence of statistical quantification makes it impossible to distinguish a true performance difference from sampling noise. Add mean trajectories with variance bands, final-error distributions, a docking success rate, and pairwise statistical comparisons or effect sizes.
  3. [Table II, Section III.B.4, and Section V.B] The study varies only payload mass and spawn radius in both training DR (Section III.B.4) and evaluation (Table II), which constitutes a single disturbance channel. The introduction and conclusion motivate the work through 'dynamic and uncertain environments' including currents, waves, limited visibility, and sensor noise, but none of these appear in the evaluation. This is not merely a scope issue: the claim that a naive policy is effective under 'realistic payloads' depends on the mass-distribution shift being a representative proxy for the sim2real gap, and the paper provides no evidence for that proxy. The claims in Section V and Section VI should be narrowed to mass-distribution shifts, or additional disturbance types should be included.
  4. [Section III.B.3 and Equations (1)-(3)] The reward function weights lambda1=0.2 and lambda2=0.03 are hand-tuned with no sensitivity analysis, and the paper never defines a quantized docking success criterion. As a result, the central statement that a naive policy is 'effective' is not operationalized, and it is unclear whether the architecture-ranking conclusions are robust to reasonable changes in the reward weights. Report the distribution of final position/orientation errors and a success rate based on explicit thresholds, and show that the main conclusions are stable across reward weight variations.
minor comments (6)
  1. [Section III.A.1] There is a typo in 'Particuular' that should read 'Particular'.
  2. [Section III.B.3] The heading 'Reward Function F ormulation' contains an unintended space; it should read 'Reward Function Formulation'.
  3. [Section I] In the second paragraph, 'and AUV will attempt to land' should be 'an AUV will attempt to land'.
  4. [Section IV.C] There is an extra space before the period at the end of the sentence '... different policies dock the AUV .'
  5. [Section III.B.5 and Experimental Setup] No hyperparameters for PPO (learning rate, clip ratio, network width/depth, episode length, discount factor) are reported, and no code or data availability statement is given, which limits reproducibility of the training runs.
  6. [Figures 4-9] The figure captions should state whether the plotted curves are means over seeds, episodes, or both, and should include error bands or shaded regions to convey variance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the policy comparison is an empirical evaluation within a fixed simulator, and the self-citation to prior simulator code is not load-bearing evidence for the robustness conclusion.

full rationale

The central claim, that a naively trained PPO docking policy is effective under simulated payload shifts and that DR/history provide only marginal gains, is an empirical result from training and evaluation runs inside one simulator. The reward function (Eq. 1-3) and the evaluation metrics are both based on position and orientation error, but this is the task definition rather than a fitted prediction: no parameter is fit to the evaluation scenarios, and the hard payload condition (7.0 kg at 0.3 m) lies outside the Large DR training range ([0,5.0] kg at 0.3 m), so the robustness result is not forced by construction. The paper's use of Cai et al. [5] is a self-citation that supplies the simulator substrate and dynamics model, but the paper does not rely on [5] to prove the robustness finding; the cross-architecture comparison is generated by the present experiments. The acknowledged simulation-only limitation (Section III.C) is an external-validity caveat, not circularity.

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

The paper's contribution is empirical and rests on simulator fidelity and task design choices. The main free parameters are the reward weights and the training distribution settings; the main axioms are the simulator's hydrodynamic model, the clean-observation assumption, and the assumption that simulated payload shifts represent the sim2real gap.

free parameters (3)
  • Reward weights λ1, λ2 = λ1=0.2, λ2=0.03
    Hand-tuned by starting with λ2=0 and increasing until trajectories appeared stable (Section III.B.3).
  • Large DR payload mass range and spawn radius = [0, 5.0] kg, 0.3 m
    Experimental condition chosen by the authors; defines the training distribution that the 'large DR' policies see.
  • History length h = 3
    Chosen for the history-conditioned policy; the paper does not justify why 3 rather than other values.
assumptions (4)
  • domain assumption The simulator's hydrodynamic model (MuJoCo simple inertial box model plus PhysX collisions, zero-latency thrusters, Yoerger thruster model) accurately captures the dynamics relevant to docking.
    The paper states the dynamics model in Section III.A.2; if the model is wrong, the robustness findings may not transfer to physical AUVs.
  • domain assumption The policy has access to clean pose and twist estimates (position, quaternion, velocities) without perception noise.
    Observations are defined in Section III.B.2; the conclusion explicitly conditions its recommendation on 'given pose and twist estimates', so sensor noise and AprilTag localization errors are excluded.
  • ad hoc to paper Adding payload mass at a point around the vehicle is a representative proxy for real-world payload configuration changes and the sim2real gap.
    The paper uses this proxy in Section III.B.4 and III.C without real-world validation; this is the weakest load-bearing premise.
  • domain assumption Thruster forces are converted using Yoerger et al. model with no latency.
    Stated in III.A.2; latency and thruster dynamics are known to affect underwater vehicle control, and ignoring them may overstate policy robustness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to Dock: A Simulation-based Study on Closing the Sim2Real Gap in Autonomous Underwater Docking." pith.science (2026). https://pith.science/paper/3Z4VADGD

@misc{pith2026250617823,
  author       = {Pith},
  title        = {Pith review of: Learning to Dock: A Simulation-based Study on Closing the Sim2Real Gap in Autonomous Underwater Docking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3Z4VADGD}},
  note         = {Machine review of arXiv:2506.17823}
}
read the original abstract

Autonomous Underwater Vehicle (AUV) docking in dynamic and uncertain environments is a critical challenge for underwater robotics. Reinforcement learning is a promising method for developing robust controllers, but the disparity between training simulations and the real world, or the sim2real gap, often leads to a significant deterioration in performance. In this work, we perform a simulation study on reducing the sim2real gap in autonomous docking through training various controllers and then evaluating them under realistic disturbances. In particular, we focus on the real-world challenge of docking under different payloads that are potentially outside the original training distribution. We explore existing methods for improving robustness including randomization techniques and history-conditioned controllers. Our findings provide insights into mitigating the sim2real gap when training docking controllers. Furthermore, our work indicates areas of future research that may be beneficial to the marine robotics community.

Figures

Figures reproduced from arXiv: 2506.17823 by the authors.

Figure 1
Figure 1. Visualization of docking trajectory in simulation environment. The vehicle starts at a random position in front of the docking station. At every iteration, the vehicle’s state is passed into a trained policy which outputs thruster commands, and then those commands are passed into the simulation to update the vehicle’s state. The goal is for the vehicle to land inside the docking station, minimizing its position to i… view at source ↗
Figure 2
Figure 2. Our general flow for training and evaluating policies for au￾tonomous docking. We utilize environments built with the Isaac Sim framework to both train and evaluate our policies. and then evaluate them under simulation but with different payloads to test robustness. To standardize our results as much as possible, we specifically develop our controller for the BlueROV2 Heavy, though our methods are easily adaptable t… view at source ↗
Figure 3
Figure 3. Demonstration of Isaac Sim’s capability for highly-parallelized training. hydrodynamics are modeled based on MuJoCo’s simple inertial box model [18]. For simplicity, thrusters are assumed to have no latency and are modeled using a zero-order dynamics model. Then, the angular velocities of the thrusters are converted to linear forces using the model proposed by Yoerger et al. [19]. 3) Docking Environment: For the doc… view at source ↗
Figures from the paper (5 more)
Figure 6
Figure 6. Figure 6: We hypothesize that the similarity in performance [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 4
Figure 4. Figure 4: Positional error over time under the easy payload configuration. Angular Error: With regards to angular error, we see fairly significant differences between different policies, as can be seen in Figures 7, 8, 9. We do observe across all payload settings, the policies t…
Figure 5
Figure 5. Figure 5: Positional error over time under the medium payload configuration [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Positional error over time under the hard payload configuration. as DR should slightly worsen performance by forcing the policy to be robust across a wider range of conditions, but incorporating history should enable the policy to condition its behavior on different en…
Figure 9
Figure 9. Figure 9: Angular error over time under the hard payload configuration [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Marinarium: A Modular Experimental Facility for Reproducible Maritime and Space-Analog Field Robotics

    cs.RO 2026-02 conditional novelty 6.0 of 10

    A modular indoor water tank with dual motion capture, a digital twin, and a space-lab link demonstrates four experiments bridging maritime and space field robotics.

Reference graph

Works this paper leans on

22 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Auv-assisted subsea exploration method in 6g enabled deep ocean based on a cooperative pac-men mechanism,

    J. Zhang, G. Han, J. Sha, Y . Qian, and J. Liu, “Auv-assisted subsea exploration method in 6g enabled deep ocean based on a cooperative pac-men mechanism,”Trans. Intell. Transport. Sys., vol. 23, no. 2, p. 1649–1660, Feb. 2022. [Online]. Available: https://doi.org/10.1109/TITS.2021.3102995

    arXiv 2022

  2. [2]

    Discov- ering biological hotspots with a passively listening auv,

    S. McCammon, S. Jamieson, T. A. Mooney, and Y . Girdhar, “Discov- ering biological hotspots with a passively listening auv,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 3789–3795

    work page 2024

  3. [3]

    An architecture for using autonomous underwater vehicles in wireless sensor networks for underwater pipeline monitoring,

    I. Jawhar, N. Mohamed, J. Al-Jaroodi, and S. Zhang, “An architecture for using autonomous underwater vehicles in wireless sensor networks for underwater pipeline monitoring,”IEEE Trans. Ind. Informatics, vol. 15, no. 3, pp. 1329–1340, 2019. [Online]. Available: https://doi.org/10.1109/TII.2018.2848290

    arXiv 2019

  4. [4]

    Sim-to-real: Learning agile locomotion for quadruped robots,

    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,”CoRR, vol. abs/1804.10332, 2018. [Online]. Available: http://arxiv.org/abs/1804.10332

    arXiv 2018

  5. [5]

    Learning to swim: Reinforcement learning for 6-dof control of thruster-driven autonomous underwater vehicles,

    L. Cai, K. Chang, and Y . Girdhar, “Learning to swim: Reinforcement learning for 6-dof control of thruster-driven autonomous underwater vehicles,” 2025, to appear. [Online]. Available: https://arxiv.org/abs/2410.00120

    arXiv 2025

  6. [6]

    Docking control of an autonomous underwater vehicle using reinforcement learning,

    E. Anderlini, G. G. Parker, and G. Thomas, “Docking control of an autonomous underwater vehicle using reinforcement learning,” Applied Sciences, vol. 9, no. 17, 2019. [Online]. Available: https://www.mdpi.com/2076-3417/9/17/3456

    work page 2019

  7. [7]

    Autonomous underwater docking using flow state estimation and model predictive control,

    R. Vivekanandan, D. Chang, and G. A. Hollinger, “Autonomous underwater docking using flow state estimation and model predictive control,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 1062–1068

    work page 2023

  8. [8]

    Deep Reinforcement Learning for Continuous Docking Control of Autonomous Underwater Vehicles: A Benchmarking Study

    M. Patil, B. Wehbe, and M. Valdenegro-Toro, “Deep reinforcement learning for continuous docking control of autonomous underwater vehicles: A benchmarking study,”CoRR, vol. abs/2108.02665, 2021. [Online]. Available: https://arxiv.org/abs/2108.02665

    work page Pith review arXiv 2021

Show all 22 references
  1. [9]

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

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

    2022

  2. [10]

    Sim-to-(multi)-real: Transfer of low-level robust control policies to multiple quadrotors,

    A. Molchanov, T. Chen, W. H ¨onig, J. A. Preiss, N. Ayanian, and G. S. Sukhatme, “Sim-to-(multi)-real: Transfer of low-level robust control policies to multiple quadrotors,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 59–66

    2019

  3. [11]

    Sim-to-real learning for humanoid box loco-manipulation,

    J. Dao, H. Duan, and A. Fern, “Sim-to-real learning for humanoid box loco-manipulation,” 2023. [Online]. Available: https://arxiv.org/abs/2310.03191

    2023 arXiv

  4. [12]

    Blind bipedal stair traversal via sim-to-real reinforcement learning,

    J. Siekmann, K. Green, J. Warila, A. Fern, and J. W. Hurst, “Blind bipedal stair traversal via sim-to-real reinforcement learning,”CoRR, vol. abs/2105.08328, 2021. [Online]. Available: https://arxiv.org/abs/2105.08328

    2021 arXiv

  5. [13]

    Learning memory-based control for human-scale bipedal locomotion,

    J. Siekmann, S. Valluri, J. Dao, L. Bermillo, H. Duan, A. Fern, and J. W. Hurst, “Learning memory-based control for human-scale bipedal locomotion,”CoRR, vol. abs/2006.02402, 2020. [Online]. Available: https://arxiv.org/abs/2006.02402

    2006 arXiv

  6. [14]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automa- tion Letter...

    2023

  7. [15]

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

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac gym: High performance gpu-based physics simulation for robot learning,”CoRR, vol. abs/2108.10470, 2021. [Online]. Available: https://arxiv.o...

    2021 arXiv

  8. [16]

    Aerial gym simulator: A framework for highly parallelized simulation of aerial robots,

    M. Kulkarni, W. Rehberg, and K. Alexis, “Aerial gym simulator: A framework for highly parallelized simulation of aerial robots,”IEEE Robotics and Automation Letters, vol. 10, no. 4, pp. 4093–4100, 2025

    2025

  9. [17]

    CUREE: A curious un- derwater robot for ecosystem exploration,

    Y . Girdhar, N. McGuire, L. Cai, S. Jamieson, S. McCammon, B. Claus, J. E. S. Soucie, J. E. Todd, and T. A. Mooney, “CUREE: A curious un- derwater robot for ecosystem exploration,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 11 411– 11 417

    2023

  10. [18]

    Mujoco: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “Mujoco: A physics engine for model-based control,” in2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 5026–5033

    2012

  11. [19]

    The influence of thruster dynamics on underwater vehicle behavior and their incorporation into control system design,

    D. Yoerger, J. Cooke, and J.-J. Slotine, “The influence of thruster dynamics on underwater vehicle behavior and their incorporation into control system design,”IEEE Journal of Oceanic Engineering, vol. 15, no. 3, pp. 167–178, 1990

    1990

  12. [20]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” CoRR, vol. abs/1707.06347, 2017. [Online]. Available: http://arxiv.org/abs/1707.06347

    2017 arXiv

  13. [21]

    Learning agile robotic locomotion skills by imitating animals,

    X. B. Peng, E. Coumans, T. Zhang, T. E. Lee, J. Tan, and S. Levine, “Learning agile robotic locomotion skills by imitating animals,”CoRR, vol. abs/2004.00784, 2020. [Online]. Available: https://arxiv.org/abs/2004.00784

    2004 arXiv

  14. [22]

    Bootstrap- ping reinforcement learning with imitation for vision-based agile flight,

    J. Xing, A. Romero, L. Bauersfeld, and D. Scaramuzza, “Bootstrap- ping reinforcement learning with imitation for vision-based agile flight,” 2024. [Online]. Available: https://arxiv.org/abs/2403.12203

    2024 arXiv

Pith tools

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