Pith. sign in

REVIEW 4 major objections 6 minor 32 references

Shared Control of Holonomic Wheelchairs through Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that reinforcement learning can provide shared control for a holonomic wheelchair by converting a simple 2D joystick input into a safe 3D motion that follows the driver's intent, keeps the wheelchair oriented toward the…

desk verdict Credible sim-to-real RL demo for holonomic wheelchair shared control, but unit-norm joystick training leaves speed authority out-of-distribution and the safety claim overreaches. read the letter →

arxiv 2507.17055 v1 pith:2DGVTFAE submitted 2025-07-22 cs.RO cs.LG

Jannis Bähler , Diego Paez-Granados , Jorge Peña-Queralta This is my paper
classification cs.ROcs.LG
keywords reinforcementlearningsharedcontrolholonomicwheelchairomnidirectionalmobilityLiDARnavigationheadingmetricsim-to-realtransfercognitiveload
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 aims to establish that reinforcement learning can make shared control of a holonomic (omnidirectional) wheelchair safe, comfortable, and intuitive: a policy reads a 2D joystick input plus 360° LiDAR and outputs a full 3D motion $(v_x, v_y, \omega)$, so the wheelchair follows the driver's intention in open space and takes over lateral and rotational degrees of freedom only when needed to avoid obstacles. The authors argue this reduces cognitive load through a heading metric $\Phi$, the angle between the chair's orientation and the driver's intended direction, and that the learned controller is smoother than or competitive with a non-learning velocity-obstacle baseline. They train in simulated environments with PPO, evaluate in Gazebo, and transfer the policy to a real omnidirectional wheelchair, reporting this as the first real-world RL-based shared control for such a platform.

What carries the argument

The central mechanism is a PPO-trained neural policy that maps the observation $(d_1,\ldots,d_n,u_x,u_y,v_x,v_y,\omega,a_{t-1},a_{t-2})$ to a velocity action $(v_x,v_y,\omega)$ for the wheelchair, with a reward function that separates obstacle avoidance, heading, and smoothness. The load-bearing design choice is the heading term $\Phi$, the angle between the wheelchair's yaw and the direction toward the unobserved target implied by the joystick, because it translates the unobservable POMDP quantity "user intention" into a measurable training signal. A curriculum over empty, cylinder, box, and door environments, plus action-smoothing terms, is what lets the resulting policy behave competently in tight spaces while keeping trajectories smooth.

What would settle it

Have a user drive the real wheelchair with a manual joystick through a long, tight corridor about 1.5 m wide and several metres long; the paper's own manual trials show this is the edge case where the policy stalls or retracts, and if this failure reproduces consistently, the claimed collision-free, intent-following shared control beyond training-like setups is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that an end-to-end policy trained with PPO can solve the shared-control problem for 3-DOF wheelchairs without knowing the user's goal: the target is never given to the agent, only the joystick direction and LiDAR. Comparing architectures and reward variants, the best policy uses a small LiDAR CNN and an LSTM, and a reward that tracks forward velocity while penalising lateral motion and heading error above a threshold; it passes a 1 m door in simulation, keeps heading smaller than the reactive baseline RDS, and shows lower or comparable jerk. Real-robot runs around a box and through a tight gap reproduce the simulated behavior, supporting the paper's claim that RL shared control transfers to real holonomic platforms.

Load-bearing premise

The system assumes a driver always points the joystick straight at the destination at full speed; real drivers correct course, oscillate, and vary joystick pressure, so the learned link between joystick direction and intention can break.

Editorial extensions

If this is right

  • A driver only needs to provide a 2D joystick direction; the policy resolves the remaining rotational and lateral degrees of freedom, which should reduce the number of corrective joystick movements, the paper's proxy for cognitive load.
  • The same training setup can be reused for any holonomic base by swapping the robot model in the simulator, because the observation and action spaces are platform-generic.
  • Reward design choices are consequential: method 2 (track only forward velocity, punish lateral motion) is smoother and is the only variant able to pass a 1 m door, suggesting that minimal lateral assistance interferes less with user authority.
  • The policy can be deployed on real hardware without retraining, and its behavior in the real-world experiments matched the simulated behavior.
  • The collision model can be capsule, circle, or square, so the approach extends to asymmetric or non-convex robot shapes rather than being tied to one footprint.

Reading between the lines

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

  • Editorial inference: if $\Phi$ is accepted as a proxy for cognitive load, then any shared-control method, RL or reactive, could be evaluated on the same axis; the paper only compares against one reactive baseline, so the claim is competitiveness with RDS, not general superiority.
  • Editorial inference: the failure mode in an unseen long corridor (one of five manual trials reached the target) indicates that the bottleneck is training diversity, not the RL formulation, so adding domain randomization, moving obstacles, or an imitation-based user model could extend coverage.
  • Editorial inference: the training user model, in which the joystick is always a unit vector toward the target, is conservative; recording real human joystick traces and checking whether the learned mapping still minimizes $\Phi$ would be a natural next test.
  • Editorial inference: the qualitative real-world validation suggests a path toward deployment, but without participant studies the cognitive-load reduction remains a design objective rather than a measured outcome.
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 / 6 minor

Summary. The paper proposes a reinforcement learning-based shared control method for a holonomic (omnidirectional) wheelchair. The policy takes a 2D user joystick input and LiDAR readings and outputs a 3D velocity command, with the goal of tracking user intent while avoiding collisions, orienting the wheelchair toward the target, and maintaining smooth motion. The method is trained in Isaac Gym with a simulated user model and curriculum over four environments, evaluated in Gazebo against several architecture/reward variants and against the non-learning Reactive Driving Support (RDS) baseline, and finally deployed on the DAA V1 holonomic wheelchair in a qualitative real-world validation. The paper claims collision-free navigation, reduced cognitive load via a new 'heading' metric, and the first real-world implementation of RL-based shared control for an omnidirectional mobility platform.

Significance. If the claims are substantiated, the work is a useful step for shared control of holonomic mobility platforms, which have received less attention than non-holonomic wheelchairs. The paper's strengths include a clear problem formulation, the use of Isaac Gym with curriculum training, a systematic comparison of architectures and reward functions, a proposed cognitive-load metric (heading), and a genuine sim-to-real transfer demonstration. These are valuable and go beyond many simulation-only studies. However, the central claims—especially 'ensures collision-free navigation' and better/competitive smoothness versus RDS—are currently stronger than what the evaluation supports, and the training user model has a load-bearing limitation regarding joystick magnitude that must be addressed.

major comments (4)
  1. [IV-A, IV-C, Table I] The training user model normalizes the joystick input to norm one, yet at test time the input norm varies with joystick extension (Section IV-A, 'the vector for the user input is normalized to have norm one' and 'in test time the user input can have different norms'). The reward function in Table I contains no term that matches commanded speed to joystick deflection magnitude: the tracking rewards compare the unit-norm user vector (ux, uy) with the velocity, so only direction, not magnitude, is rewarded. Consequently, half-deflection and centered-joystick inputs are out-of-distribution, and a user may receive full commanded speed even when partially releasing the joystick, or sustain speed when centering it. This directly threatens the shared-control claim and the abstract's 'ensures collision-free navigation' assertion, because the user may not have effective speed authority. The authors must either train/evaluate with variable joystick magnitudes (including zero) or explicitly limit their claims to direction-only joystick control and discuss the safety implications.
  2. [V-B2, V-C, Abstract] The manual user input and real-world validations are qualitative and lack basic experimental rigor. Section V-B2 reports that for target T3 only one of five trials reached the target, and Section VI states the real experiments were qualitative, yet the abstract claims 'ensures collision-free navigation.' There is no reporting of the number of users, no inter-trial statistics, no success-rate quantification for the real-world tasks (Appendix D shows metrics but no clear success criteria), and no baseline comparison in real-world runs. To support 'ensures collision-free navigation' and 'smartly orienting the wheelchair,' the evaluation must include quantitative success rates, confidence intervals or statistical tests, and a definition of what 'ensures' means in the presence of the reported T3 failure. As written, the claim overstates the evidence.
  3. [V-B1, Figure 7] The comparison to RDS is confounded by design: Figure 7 uses a subset of the tasks, and the paper states that RDS has a 3D input while the RL method uses a 2D input. RDS is therefore evaluated in an unfair configuration—once it turns away from the target it never turns back—so the resulting 'superior heading' of the RL approach is partly a consequence of input-space mismatch rather than control quality. The sentence 'Generally, the heading performance of the RL approach is superior' is not supported as a general statement. At minimum, the comparison should either give RDS a matching 2D input interface (e.g., derive (ux, uy, ω_user) from the 2D joystick and heading error) or be presented explicitly as an illustration of the different input modalities, not as a benchmark. The current framing is load-bearing for the claim of 'better or competitive smoothness compared to a previous non-learning-based method.'
  4. [V-B, Figures 5 and 7] The comparison of models in Figure 5 and the RL-versus-RDS comparison in Figure 7 report no statistical measures: no error bars, no repeated-run variability, no confidence intervals, and no significance tests. The plots appear to show single trajectories or single aggregate values per condition, which is insufficient to conclude that one architecture or reward method is better than another, especially given the observed variation between tasks and incident angles. The paper's comparative claims (e.g., 'the usage of an LSTM generally improves performance,' 'method 2 ... jerk ... much smaller') require repeated trials and basic statistics. This is a general evaluation gap that affects the validity of the model-selection conclusion (choosing SCLFC_D_R2 as best-performing).
minor comments (6)
  1. [Title page] The affiliation contains a typo: 'Artifical Ingelligece' should be 'Artificial Intelligence' in 'Centre for Artificial Ingelligece, Zurich University of Applied Sciences - ZHAW.'
  2. [Table I] The heading threshold is written as 'Φtreash' in the table but as 'Φthresh' in the text and Table II; please unify the notation.
  3. [Section IV-C and Table II] The caption 'If there is no value for rvy method 1 is used, otherwise method 2' is confusing because Table II lists '-' for some entries; clarify that '-' means 'not used' and explicitly state that rvy is only present for method 2.
  4. [Section V-B1] The term 'fliers' is used without definition; if it means outlier spikes in jerk, define it and explain how it is computed from the jerk signal.
  5. [Section IV-B] The user input description says '(ux, uy ∈ [0, 1])', but the input should include negative values for direction; use [-1, 1]^2 or clarify the notation.
  6. [Section VI] The conclusion states 'our approach allows various collision model shapes to cover both convex and non-convex shapes'; this appears only as a short statement and is not demonstrated in the experiments, so either provide evidence or mark it as a feature of the framework rather than a validated result.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training, reward, and evaluation loop is standard, and no central claim reduces to its own inputs.

full rationale

I walked the paper's derivation chain from the user model (Section IV-A) through the reward function (Section IV-C) to the evaluation metrics (Section V-B). The policy is trained with reward terms that penalize heading and jerk, and the same quantities are later reported as performance metrics; however, the paper never presents these metrics as predictions derived from the model. They are training objectives and evaluation benchmarks, which is standard reward shaping rather than circular reasoning. The user model assumes a unit-norm joystick vector pointing at a fictional target, and the reward tracks that vector; the evaluation then measures the resulting heading against the same simulated user input. This is a self-consistent training/evaluation loop, not a derivation that presupposes its conclusion. The RDS baseline is prior work from the same laboratory, but it is used only as a comparative benchmark; the holonomic extension is described as an in-laboratory implementation, not cited as an external theorem that forces the RL result. No uniqueness claim is imported from the authors' prior work, no fitted parameter is relabeled as a prediction, and no known result is merely renamed. The admitted limitations—unit-norm user model, out-of-distribution joystick magnitudes, qualitative real-world tests, and robustness degradation in unseen environments—are validity and generalization concerns, not circularity. No load-bearing step reduces by construction to its own inputs.

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

The central claim rests on hand-tuned reward and collision parameters, a simplified simulated user model, and the unvalidated proxy that heading and jerk capture cognitive load and comfort. The sim-to-real transfer assumes Isaac Gym with Kaya matches the real wheelchair sufficiently without domain randomization. No new physical entities are introduced.

free parameters (5)
  • Reward weights (rc, rcrit, rcol, rh, Phi_thresh, ra, rl, rvy, ras) = varies by model; e.g., SCLFC_D_R2: -1, -1, -100, -0.5, 0.2, 0.5, -0.5, -1.6, -0.02
    Hand-tuned in Table II to balance obstacle avoidance, heading, tracking, and smoothness; no sensitivity analysis or principled tuning method is reported.
  • Collision model parameters (dcol, dcrit, capsule spacing) = 0.65 m diameter, 1.05 m diameter, 0.3 m spacing
    Chosen by hand (Section V-A2) to define the safety buffer around the robot; these values affect the reward landscape and collision resets.
  • Velocity scaling limits = 1 m/s, 1 rad/s in training; 0.67 m/s, 2 rad/s in Gazebo
    Maximum velocities used to scale the normalized action space (Sections IV-B, V-A2); adjusted between simulators due to tracking discrepancies.
  • Environment randomization ranges = box l in [1,4] m, b in [1,2] m; door width in [0.9,1.75] m
    Hand-selected ranges for training environments (Section IV-A); they define the distribution the policy sees.
  • PPO hyperparameters = lr 5e-4, entropy 1e-2, clip 0.2, horizon 128, minibatch 4096, 4 mini-epochs
    Standard PPO settings reported in Section V-A1; chosen for stable training, not derived from the problem.
assumptions (5)
  • domain assumption User always points the joystick toward the target during training, with a unit-norm vector
    User Model in Section IV-A. This is the foundation of the simulated user input and the definition of heading Phi; the paper notes in future work that a more complex user model is needed (Section VI).
  • domain assumption The heading Phi between robot orientation and target direction is a valid proxy for cognitive load
    Introduced in Section I and used as both reward and evaluation metric (Sections IV-A, V-B); no user study or prior literature is cited to validate this link.
  • domain assumption Velocity jerk measures user comfort
    Section V-B uses jerk as the comfort metric for evaluating smoothing; this is an accepted but unvalidated proxy, and no ablation links jerk values to user ratings.
  • domain assumption Isaac Gym with the Kaya model captures the dynamics of the real DAA V1 wheelchair sufficiently for zero-shot transfer
    Training is done on Kaya in Isaac Gym, then the policy is run on a DAA V1 real robot (Sections V-A, V-C) without domain randomization or dynamics adaptation; the paper only adjusts maximum velocity limits.
  • domain assumption PPO converges to a policy that generalizes across the four training environments
    The paper relies on PPO (Section III-B) and curriculum learning (Section V-A1) to produce a transferable policy; convergence and generalization are empirical, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Shared Control of Holonomic Wheelchairs through Reinforcement Learning." pith.science (2026). https://pith.science/paper/2DGVTFAE

@misc{pith2026250717055,
  author       = {Pith},
  title        = {Pith review of: Shared Control of Holonomic Wheelchairs through Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DGVTFAE}},
  note         = {Machine review of arXiv:2507.17055}
}
read the original abstract

Smart electric wheelchairs can improve user experience by supporting the driver with shared control. State-of-the-art work showed the potential of shared control in improving safety in navigation for non-holonomic robots. However, for holonomic systems, current approaches often lead to unintuitive behavior for the user and fail to utilize the full potential of omnidirectional driving. Therefore, we propose a reinforcement learning-based method, which takes a 2D user input and outputs a 3D motion while ensuring user comfort and reducing cognitive load on the driver. Our approach is trained in Isaac Gym and tested in simulation in Gazebo. We compare different RL agent architectures and reward functions based on metrics considering cognitive load and user comfort. We show that our method ensures collision-free navigation while smartly orienting the wheelchair and showing better or competitive smoothness compared to a previous non-learning-based method. We further perform a sim-to-real transfer and demonstrate, to the best of our knowledge, the first real-world implementation of RL-based shared control for an omnidirectional mobility platform.

Figures

Figures reproduced from arXiv: 2507.17055 by the authors.

Figure 1
Figure 1. Illustration of the holonomic wheelchair used in this study. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The user input always points towards the target, while the heading [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. General network architecture with layer names. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: We evaluate our agents based on heading and jerk (as defined in the text). The titles of the subplots refer to the task name, obstacle size and incident [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: To compare the RL approach to RDS, we use the same setup as in 5. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Trajectories (red), user input (green), and LiDAR data (white) of real [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 15 canonical work pages

  1. [1]

    A robotic wheelchair roaming in a railway station,

    E. Prassler, J. Scholz, and P. Fiorini, “A robotic wheelchair roaming in a railway station,” Dec. 1998

    work page 1998

  2. [2]

    Dynamic shared control for human- wheelchair cooperation,

    Q. Li, W. Chen, and J. Wang, “Dynamic shared control for human- wheelchair cooperation,” in 2011 IEEE International Conference on Robotics and Automation , 2011, pp. 4278–4283. DOI: 10.1109/ICRA. 2011.5980055

    arXiv 2011

  3. [3]

    Probabilistic vs linear blending approaches to shared control for wheelchair driving,

    C. Ezeh, P. Trautman, L. Devigne, V . Bureau, M. Babel, and T. Carlson, “Probabilistic vs linear blending approaches to shared control for wheelchair driving,” in 2017 International Conference on Rehabili- tation Robotics (ICORR) , 2017, pp. 835–840. DOI: 10.1109/ICORR. 2017.8009352

    work page arXiv 2017

  4. [4]

    Understanding Shared Control for Assistive Robotic Arms

    K. Kronhardt, M. Pascher, and J. Gerken, Understanding shared control for assistive robotic arms, 2023. arXiv: 2303.01993 [cs.HC]. [Online]. Available: https://arxiv.org/abs/2303.01993

    work page Pith review arXiv 2023

  5. [5]

    Reactive navigation in crowds for non-holonomic robots with convex bounding shape,

    D. J. Gonon, D. Paez-Granados, and A. Billard, “Reactive navigation in crowds for non-holonomic robots with convex bounding shape,” IEEE Robotics and Automation Letters , vol. 6, no. 3, pp. 4728–4735, 2021. DOI: 10.1109/LRA.2021.3068660

    arXiv 2021

  6. [6]

    T. Fan, X. Cheng, J. Pan, D. Manocha, and R. Yang, Crowdmove: Autonomous mapless navigation in crowded scenarios , 2018. arXiv: 1807.07870 [cs.RO]. [Online]. Available: https://arxiv.org/abs/1807. 07870

    work page Pith review arXiv 2018

  7. [7]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, pp. 1735–80, Dec. 1997. DOI: 10.1162/neco.1997. 9.8.1735

    work page doi:10.1162/neco.1997 1997

  8. [8]

    Wijmans, M

    E. Wijmans, M. Savva, I. Essa, S. Lee, A. S. Morcos, and D. Batra, Emergence of maps in the memories of blind navigation agents , 2023. arXiv: 2301.13261 [cs.AI]. [Online]. Available: https://arxiv.org/ abs/2301.13261

    arXiv 2023

Show all 32 references
  1. [9]

    Rloc: Terrain-aware legged locomotion using reinforcement learning and optimal control,

    S. Gangapurwala, M. Geisert, R. Orsolino, M. Fallon, and I. Havoutis, “Rloc: Terrain-aware legged locomotion using reinforcement learning and optimal control,” IEEE Transactions on Robotics , vol. 38, no. 5, pp. 2908–2927, 2022

    2022

  2. [10]

    Learning robust autonomous navigation and locomotion for wheeled- legged robots,

    J. Lee, M. Bjelonic, A. Reske, L. Wellhausen, T. Miki, and M. Hutter, “Learning robust autonomous navigation and locomotion for wheeled- legged robots,” Science Robotics, vol. 9, no. 89, Apr. 2024, ISSN : 2470-

    2024

  3. [11]

    Radosavovic, T

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath, Real-world humanoid locomotion with reinforcement learn- ing, 2023. arXiv: 2303.03381 [cs.RO]. [Online]. Available: https: //arxiv.org/abs/2303.03381

    2023 arXiv

  4. [12]

    Reaching the limit in autonomous racing: Optimal control versus reinforcement learning,

    Y . Song, A. Romero, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Reaching the limit in autonomous racing: Optimal control versus reinforcement learning,” Science Robotics , vol. 8, no. 82, Sep. 2023, ISSN : 2470-9476. DOI: 10 . 1126 / scirobotics . adg1462. [Online]. Availabl...

    2023 doi

  5. [13]

    Towards rl-based hydraulic excavator automa- tion,

    P. Egli and M. Hutter, “Towards rl-based hydraulic excavator automa- tion,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 2692–2697. DOI: 10.1109/IROS45743. 2020.9341598

    2020

  6. [14]

    Robot navi- gation in dense human crowds: Statistical models and experimental studies of human–robot cooperation,

    P. Trautman, J. Ma, R. M. Murray, and A. Krause, “Robot navi- gation in dense human crowds: Statistical models and experimental studies of human–robot cooperation,” The International Journal of Robotics Research , vol. 34, no. 3, pp. 335–356, 2015. DOI: 10 . 1177 / 02783649145...

    2015

  7. [15]

    C. Chen, Y . Liu, S. Kreiss, and A. Alahi, Crowd-robot interaction: Crowd-aware robot navigation with attention-based deep reinforce- ment learning , 2019. arXiv: 1809.08835 [cs.RO]. [Online]. Avail- able: https://arxiv.org/abs/1809.08835

    2019 arXiv

  8. [16]

    S. Yao1, G. Chen, Q. Qiu, J. Ma, X. Chen, and J. Ji, Crowd-aware robot navigation for pedestrians with multiple collision avoidance strategies via map-based deep reinforcement learning , 2021. arXiv: 2109.02541 [cs.RO]. [Online]. Available: https://arxiv.org/abs/2109.02541

    2021 arXiv

  9. [17]

    Walk these ways: Tuning robot control for generalization with multiplicity of behavior,

    G. B. Margolis and P. Agrawal, “Walk these ways: Tuning robot control for generalization with multiplicity of behavior,” Conference on Robot Learning, 2022

    2022

  10. [18]

    A. R. Geist, J. Frey, M. Zobro, A. Levina, and G. Martius, Learning with 3d rotations, a hitchhiker’s guide to so(3) , 2024. arXiv: 2404. 11735 [cs.LG]. [Online]. Available: https : / / arxiv. org / abs / 2404 . 11735

    2024

  11. [19]

    Motion planning in dynamic envi- ronments using velocity obstacles,

    P. Fiorini and Z. Shiller, “Motion planning in dynamic envi- ronments using velocity obstacles,” The International Journal of Robotics Research , vol. 17, no. 7, pp. 760–772, 1998. DOI: 10 . 1177 / 027836499801700706. eprint: https : / / doi . org / 10 . 1177 / 027836499801700...

    1998

  12. [20]

    Paez-Granados, Y

    D. Paez-Granados, Y . He, D. Gonon, et al. , Pedestrian-robot inter- actions on autonomous crowd navigation: Reactive control methods and evaluation metrics, 2022. arXiv: 2208.02121 [cs.RO]. [Online]. Available: https://arxiv.org/abs/2208.02121

    2022 arXiv

  13. [21]

    Reddy, A

    S. Reddy, A. D. Dragan, and S. Levine, Shared autonomy via deep reinforcement learning, 2018. arXiv: 1802.01744 [cs.LG]. [Online]. Available: https://arxiv.org/abs/1802.01744

    2018 arXiv

  14. [22]

    Reinforcement learning based user-specific shared control navigation in crowds,

    B. Zhang, C. Holloway, and T. Carlson, “Reinforcement learning based user-specific shared control navigation in crowds,” Oct. 2023, pp. 4387–4392. DOI: 10.1109/SMC53992.2023.10394139

    2023

  15. [23]

    Schaff and M

    C. Schaff and M. R. Walter, Residual policy learning for shared autonomy, 2020. arXiv: 2004.05097 [cs.RO]. [Online]. Available: https://arxiv.org/abs/2004.05097

    2020 arXiv

  16. [24]

    Generative adversarial imitation learning,

    J. Ho and S. Ermon, “Generative adversarial imitation learning,” Jun. 2016

    2016

  17. [25]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, et al. , Isaac gym: High performance gpu-based physics simulation for robot learning , 2021. arXiv: 2108.10470 [cs.RO]. [Online]. Available: https://arxiv.org/ abs/2108.10470

    2021 arXiv

  18. [26]

    Reinforcement learning and markov de- cision processes,

    M. Otterlo and M. Wiering, “Reinforcement learning and markov de- cision processes,” Reinforcement Learning: State of the Art , pp. 3–42, Jan. 2012. DOI: 10.1007/978-3-642-27645-3 1

    2012 doi

  19. [27]

    Partially observable markov decision processes (pomdps) and robotics,

    H. Kurniawati, “Partially observable markov decision processes (pomdps) and robotics,” CoRR, vol. abs/2107.07599, 2021. arXiv: 2107.07599. [Online]. Available: https://arxiv.org/abs/2107.07599

    2021 arXiv

  20. [28]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, Proximal policy optimization algorithms , 2017. arXiv: 1707 . 06347 [cs.LG]. [Online]. Available: https://arxiv.org/abs/1707.06347

    2017 arXiv

  21. [29]

    Taheri and S

    H. Taheri and S. R. Hosseini, Deep reinforcement learning with enhanced ppo for safe mobile robot navigation , 2024. arXiv: 2405. 16266 [cs.RO]. [Online]. Available: https : / / arxiv. org / abs / 2405 . 16266

    2024

  22. [30]

    Engstrom, A

    L. Engstrom, A. Ilyas, S. Santurkar, et al., Implementation matters in deep policy gradients: A case study on ppo and trpo , 2020. arXiv: 2005.12729 [cs.LG]. [Online]. Available: https://arxiv.org/abs/2005. 12729

    2020 arXiv

  23. [31]

    Y . Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel, Bench- marking deep reinforcement learning for continuous control , 2016. arXiv: 1604.06778 [cs.LG]. [Online]. Available: https://arxiv.org/ abs/1604.06778. EOD EOD APPENDIX : F URTHER RESULTS A. Trajectories of simu...

    2016 arXiv

  24. [9476]

    1126 / scirobotics

    DOI: 10 . 1126 / scirobotics . adi9641. [Online]. Available: http : //dx.doi.org/10.1126/scirobotics.adi9641

    doi

Pith tools

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