Pith. sign in

REVIEW 5 major objections 5 minor 42 references

End-to-End Humanoid Robot Safe and Comfortable Locomotion Policy

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper presents an end-to-end policy that maps raw LiDAR point clouds to humanoid motor commands, with safety enforced through a Control Barrier Function-derived cost inside a constrained Markov Decision Process, and reports safer, more

desk verdict Incremental but credible humanoid locomotion policy with a real G1 demo; the central safety contribution is not isolated as written because the CBF cost is missing from Table I. read the letter →

arxiv 2508.07611 v1 pith:SJZIO2BA submitted 2025-08-11 cs.RO

Zifan Wang , Xun Yang , Jianzhuang Zhao , Jiaming Zhou , Teli Ma , Ziyao Gao , Arash Ajoudani , Junwei Liang This is my paper
classification cs.RO
keywords humanoidlocomotionLiDARperceptioncontrolbarrierfunctionsconstrainedreinforcementlearningP3Ocollisionavoidancehumancomfortsim-to-real
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a humanoid robot can navigate cluttered 3D scenes safely and comfortably with an end-to-end policy that takes raw LiDAR point clouds plus joint and command history and outputs motor commands. The safety mechanism is a translation of Control Barrier Function conditions into a cost within a Constrained Markov Decision Process, so a model-free constrained policy optimizer (P3O) can learn to respect a dynamics-aware barrier without filtering actions online. The authors argue this is more reliable than reward shaping and less reactive than a plain distance-based constraint, and they report quantitative improvements in simulation and a successful transfer to a physical humanoid in static and dynamic obstacle tests.

What carries the argument

The Linear Discrete-Time Control Barrier Function (LDCBF). For a linear model $x_{k+1} = A_L x_k + B_L u_k$ and a local hyperplane obstacle approximation, the barrier condition becomes an affine constraint on the control input. Instead of filtering actions with this constraint, the paper turns its violation into the cost $C_D(s_k,u_k) = \max\{0, -G_D(s_k,u_k)\}$ inside a CMDP solved by P3O. This lets a model-free policy learn dynamics-aware safety.

What would settle it

Measure on the physical robot the one-step-ahead signed distance predicted by the linear model against the actual signed distance after executing the commanded action. If there is a systematic mismatch—for example, the predicted safety margin $h_D$ increases while the robot actually approaches the obstacle—the cost function is enforcing a barrier that does not correspond to the real dynamics, and the claimed safety guarantee would not hold.

Watch

Extended reading notes

Core claim

The central claim is that a discrete-time CBF condition, when written for a locally linear model of the robot and a locally planar obstacle boundary, can be converted into an instantaneous cost that a model-free constrained RL algorithm can minimize. The barrier function $h_D(s_k)$ is the signed distance to the obstacle minus a margin; the one-step-ahead condition $h_D(s_{k+1}) \ge (1-\gamma_{\text{CBF}}) h_D(s_k)$ becomes an affine inequality $G_D(s_k,u_k) \ge 0$, and the policy penalizes $C_D = \max\{0, -G_D(s_k,u_k)\}$. The paper reports that training with this cost, together with comfort-oriented rewards for proxemic distance, approach velocity, and tangential avoidance, yields a policy

Load-bearing premise

The safety cost is built on the assumptions that the robot's one-step dynamics are linear ($x_{k+1}=A_L x_k + B_L u_k$) and that the nearest obstacle boundary can be treated as a flat plane; if either approximation is wrong, the cost will penalize the wrong actions.

Editorial extensions

If this is right

  • If the approach holds, safety can be enforced in model-free RL by converting model-based barrier conditions into costs, without online optimization or action filtering.
  • LiDAR-based end-to-end policies can handle obstacles that are invisible to 2D elevation maps, such as overhanging platforms.
  • Comfort rewards derived from proxemics and approach direction reduce time spent in close proximity to obstacles and improve success in confined spaces.
  • The learned policy transfers from simulation to a real humanoid, handling both static clutter and a sudden human approach.

Reading between the lines

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

  • The safety guarantee is conditional on the quality of the linear model; a natural test is to compare the one-step-ahead prediction of the barrier with actual measurements and measure how much violation of the assumption changes the policy's behavior.
  • The same CBF-to-cost translation could be applied to other constraints (e.g., joint limits, self-collision) and to other constrained RL algorithms, making the mechanism independent of P3O.
  • The paper defines comfort formally but measures it by distance thresholds; a direct user study with humans rating perceived comfort would be a stronger test of the HRI claim.
  • A comparison against an online CBF filter that projects actions into the safe set would clarify whether the learned cost is necessary or whether the same barrier could be enforced post-hoc.
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

5 major / 5 minor

Summary. The paper proposes an end-to-end humanoid locomotion policy that maps raw LiDAR point clouds, processed by a GRU encoder, together with proprioceptive and command history directly to motor commands. Safety is framed as a CMDP, and the authors' central methodological contribution is a translation of discrete-time CBF safety conditions into an instantaneous cost C_D (Eq. 7), which is then used inside the model-free P3O algorithm alongside comfort-oriented rewards. The method is evaluated against PPO-RewardShaping and P3O baselines in four Isaac Sim scenarios and deployed on a Unitree G1 humanoid in two qualitative real-world demonstrations.

Significance. If fully substantiated, the paper would provide a practical way to inject model-based CBF safety into a model-free constrained RL loop, while also demonstrating raw-LiDAR end-to-end humanoid navigation. The combination of LiDAR perception, CMDP safety, and HRI-inspired comfort rewards is timely and potentially useful. However, as currently written, the central attribution of safety improvements to the CBF-derived cost is not supported: the cost does not appear in the implementation table, the baseline is confounded, the linear-dynamics assumption underlying C_D is unvalidated, and the quantitative evidence base is thin. These are fixable with additional experiments and clarification, so the contribution remains plausible rather than established.

major comments (5)
  1. [Sec. IV.B, IV.C, Table I] The central safety mechanism C_D from Eq. (7) does not appear in the implementation. Sec. IV.C.2 lists only the binary Safety Distance Violation, joint-limit, and self-collision costs in Table I, and the P3O objective Eq. (8) is written for a generic set {C_j}. If C_D is part of the cost set, it must be named, given a threshold, and included in the training-loss description; otherwise the paper's key contribution is never trained. As written, the claim that CBF-derived costs enforce safety is unsupported.
  2. [Sec. III.B / IV.B, Eq. (3), (5)-(7)] Eq. (3) assumes linear dynamics x_{k+1}=A_L x_k + B_L u_k, and Eq. (6) uses this model to derive G_D. No identification procedure, numerical values, or validation of A_L and B_L is provided, nor an error analysis showing that the one-step-ahead signed-distance prediction is accurate enough. If the linear model is inaccurate, C_D penalizes the wrong actions and the claimed provable-safety link is broken. The authors should justify or replace this assumption, or at least report the one-step-ahead prediction error in simulation.
  3. [Sec. V.B, Tables II and III] Table II reports aggregate times from 10 runs and Table III success rates from 30 trials per scenario. No confidence intervals or significance tests are given. Differences such as 90% vs 83% in scenario (a) and 100% vs 100% in scenario (c) cannot support comparative claims. The reported improvements in Narrow Passage (33% vs 60%) and Dynamic Agents (70% vs 86%) should be accompanied by binomial confidence intervals or repeated-seed results before conclusions are drawn.
  4. [Sec. V.B.1 / Table I] The comfort evaluation measures time spent in the 0.6-1.2 m distance band, while the Proxemic Comfort reward is exp(-alpha_p(d_human - 1.2)^2) and the Safety Distance Violation cost is triggered below 0.8 m. The evaluation metric is therefore closely aligned with the trained objective. This does not invalidate the result, but it means the 'comfort' claim is not independently measured. Additional external metrics (e.g., approach-angle distributions, human ratings, or velocity variability) would strengthen the claim.
  5. [Sec. V.D] The hardware validation is only qualitative: two scenarios are described narratively and shown in figures. No number of trials, success counts, or measured safety/comfort variables are reported. Given the paper's sim-to-real claim, at least a small quantitative table (e.g., number of runs, collisions, completion time) is needed.
minor comments (5)
  1. [Throughout] Typos: 'polcy' in Sec. IV.A.1; 'ad the predictability' in Sec. IV.C.1; 'EE)' in the Fig. 6 caption; and an orphan dash in the abstract ('P3O—to enforce').
  2. [Table I and Eq. (1)] The threshold notation is inconsistent: Eq. (1) uses epsilon_j, while Table I uses d_j. Clarify that d_safe=0.8 m is the indicator threshold for the binary cost and d_j=0.0 is the constraint bound on expected cost.
  3. [Definition 2 / Sec. V.B.1] Definition 2 defines a comfortable policy as remaining within S_IC, but S_IC is never made concrete. Table II labels the 0.6-1.2 m band as 'Uncomfortable Space But Safe', which should be reconciled with the formal definition of the comfortable subspace.
  4. [Sec. V.B] The P3O baseline is described inconsistently: first as identical to P3O-CBF but without the comfort reward, and later as having a cost function 'only based on distance.' This should be disambiguated, since it directly affects the interpretation of the ablation.
  5. [References] Reference [41] lists 'G. Authors' as the author; the actual author list should be supplied. In addition, the 64-dimensional LiDAR feature extraction is not detailed; please specify the architecture or cite the extraction module.

Circularity Check

1 steps flagged · score 6.0 of 10

The comfort evaluation metric is the same distance objective as the Proxemic Comfort reward, making the comfort claim self-confirming; the safety claim retains independent content but the CBF cost is not cleanly isolated in the reported objective.

  1. self definitional [Sec. III.A (Definitions 1–2); Sec. IV.C.1 (Table I and reward text); Sec. V.B (Table II)]
    "Comfortable ... for any initial state s0 ∈ SIC , the resulting trajectory satisfies P(st ∈ SIC , ∀t ≥ 0 | s0, π) = 1 . The specific definition of SIC is based on proximity, velocity, and agent’s state. ... Proxemic Comfort exp(−αp(dhuman,k − dsocial)2) 1.5 ... dsocial is the ideal social distance of 1.2 meters. ... Time in Uncomfortable Space (0.6≤Dobs<1.2m)."

    Definition 2 makes 'comfortable' mean never leaving the proximity-based set S_IC, and Definition 1 says S_IC is 'based on proximity.' The Proxemic Comfort reward is a Gaussian centered at d_social=1.2 m, i.e., it trains the policy to keep D_obs near 1.2 m. Table II defines the 'Uncomfortable Space' as 0.6≤D_obs<1.2 m, exactly the close side of that trained target. Thus the reported comfort advantage is the policy's own reward function written as an evaluation metric; it is true by the paper's definitions and reward choice, not by an independent measure of human comfort (no user study or external comfort protocol is given). The comfort claim therefore reduces by construction to the training objective.

full rationale

One step is genuinely circular: the comfort evaluation. The paper defines comfortable policies as those that remain inside the proximity-based S_IC, trains with a Proxemic Comfort reward peaked at d_social=1.2 m, and then measures comfort as time spent with D_obs<1.2 m. The reported comfort benefit is thus an output of the reward, not an independent finding. The safety results, in contrast, have independent content: the policy is constrained at d_safe=0.8 m (and/or D_min) while the unsafe metric uses a stricter 0.6 m threshold, and success rates in Table III are task-level outcomes (reaching goals, not colliding) that are not identical to the training costs. However, the paper's headline CBF-to-cost mechanism is not cleanly isolated. Eq. (7) defines C_D, but Table I's cost list contains only the binary C_safe, joint-limit, and self-collision costs, and the P3O baseline is described both as 'same as P3O-CBF but trained without Comfort-Oriented reward' and as having a cost 'only based on distance.' This means the improvement attributed to the dynamics-aware CBF cost is confounded with the comfort-reward ablation. That is a missing-support/confound rather than a definitional circularity, so it does not by itself raise the score, but it makes the central safety attribution weaker. Overall: partial circularity in one of the two headline claims (comfort), with the safety claim still independent, yields a score of 6.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim depends on hand-set safety and comfort thresholds, unreported CBF parameters, an unvalidated linear dynamics model, and the hyperplane obstacle approximation. No new physical entities are introduced. The free parameters and domain assumptions are the main cost of the contribution because they are not derived or independently measured.

free parameters (6)
  • d_safe = 0.8 m
    Hard safety threshold used in the binary constraint cost in Table I; hand-set and defines the unsafe space.
  • d_social = 1.2 m
    Ideal proxemic distance in the comfort reward, chosen from HRI literature but not measured in this study.
  • D_min and gamma_CBF = unspecified
    Parameters in Eqs. (4)-(5) that determine the CBF safety cost; values are not reported, though they control the central safety mechanism.
  • alpha_v, alpha_omega, alpha_p = unspecified
    Exponential gains inside the velocity, yaw, and proxemic rewards in Table I; never given numerical values.
  • Reward weights in Table I = see Table I (e.g., 2.0, 0.5, -0.02, -1e-6, -2.5e-7, -5e-3, -1e-5)
    All task, auxiliary, and comfort reward component weights are hand-tuned and central to the learned behavior.
  • P3O penalty multipliers kappa_j = unspecified
    Eq. (8) uses per-constraint penalty weights kappa_j; these are not reported, so the constraint enforcement strength is unknown.
assumptions (3)
  • domain assumption The robot's dynamics are linear: x_{k+1}=A_L x_k+B_L u_k (Eq. 3)
    Used to construct the LDCBF safety condition, but humanoid dynamics are highly nonlinear and no model-identification or error analysis is provided in Section III.B.
  • domain assumption Obstacle boundaries are locally hyperplanes, so the signed distance h_D in Eq. (5) is a valid barrier function
    Needed to turn raw point clouds into a linear safety constraint; inaccurate near corners, non-convex objects, and irregular human shapes.
  • domain assumption The 64-dimensional LiDAR embedding preserves enough 3D obstacle geometry for closed-loop navigation
    The exteroceptive perception is central, but the embedding construction is not described and no analysis validates that it retains sufficient information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of End-to-End Humanoid Robot Safe and Comfortable Locomotion Policy." pith.science (2026). https://pith.science/paper/SJZIO2BA

@misc{pith2026250807611,
  author       = {Pith},
  title        = {Pith review of: End-to-End Humanoid Robot Safe and Comfortable Locomotion Policy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJZIO2BA}},
  note         = {Machine review of arXiv:2508.07611}
}
read the original abstract

The deployment of humanoid robots in unstructured, human-centric environments requires navigation capabilities that extend beyond simple locomotion to include robust perception, provable safety, and socially aware behavior. Current reinforcement learning approaches are often limited by blind controllers that lack environmental awareness or by vision-based systems that fail to perceive complex 3D obstacles. In this work, we present an end-to-end locomotion policy that directly maps raw, spatio-temporal LiDAR point clouds to motor commands, enabling robust navigation in cluttered dynamic scenes. We formulate the control problem as a Constrained Markov Decision Process (CMDP) to formally separate safety from task objectives. Our key contribution is a novel methodology that translates the principles of Control Barrier Functions (CBFs) into costs within the CMDP, allowing a model-free Penalized Proximal Policy Optimization (P3O) to enforce safety constraints during training. Furthermore, we introduce a set of comfort-oriented rewards, grounded in human-robot interaction research, to promote motions that are smooth, predictable, and less intrusive. We demonstrate the efficacy of our framework through a successful sim-to-real transfer to a physical humanoid robot, which exhibits agile and safe navigation around both static and dynamic 3D obstacles.

Figures

Figures reproduced from arXiv: 2508.07611 by the authors.

Figure 1
Figure 1. Illustration of safe and comfortable space [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed training framework. Raw sensor data is [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of trajectories from the ablation study in a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of the diverse and challenging evaluation scenarios [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Demonstration of reactive avoidance to a sudden human approach. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages

  1. [1]

    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

    work page 2022

  2. [2]

    Rapid locomotion via reinforcement learning,

    G. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal, “Rapid locomotion via reinforcement learning,” in Robotics: Science and Systems, 2022

    work page 2022

  3. [3]

    Learning agile and dynamic motor skills for legged robots. sci,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots. sci,” Robotics, vol. 4, p. 26, 2019

    work page 2019

  4. [4]

    Extreme parkour with legged robots,

    X. Cheng, K. Shi, A. Agarwal, and D. Pathak, “Extreme parkour with legged robots,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 11 443–11 450

    work page 2024

  5. [5]

    Humanoid parkour learning,

    Z. Zhuang, S. Yao, and H. Zhao, “Humanoid parkour learning,” arXiv preprint arXiv:2406.10759, 2024

    arXiv 2024

  6. [6]

    Robot parkour learning,

    Z. Zhuang, Z. Fu, J. Wang, C. Atkeson, S. Schwertfeger, C. Finn, and H. Zhao, “Robot parkour learning,” arXiv preprint arXiv:2309.05665 , 2023

    arXiv 2023

  7. [7]

    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, p. eadi9641, 2024

    2024

  8. [8]

    Depth camera technology comparison and performance evaluation,

    B. Langmann, K. Hartmann, and O. Loffeld, “Depth camera technology comparison and performance evaluation,” in International Conference on Pattern Recognition Applications and Methods , vol. 2. SciTePress, 2012, pp. 438–444

    work page 2012

Show all 42 references
  1. [9]

    Robust vision-based relative- localization approach using an rgb-depth camera and lidar sensor fusion,

    H. Song, W. Choi, and H. Kim, “Robust vision-based relative- localization approach using an rgb-depth camera and lidar sensor fusion,” IEEE Transactions on Industrial Electronics , vol. 63, no. 6, pp. 3725–3736, 2016

    2016

  2. [10]

    A survey of lidar and camera fusion enhancement,

    H. Zhong, H. Wang, Z. Wu, C. Zhang, Y . Zheng, and T. Tang, “A survey of lidar and camera fusion enhancement,” Procedia Computer Science, vol. 183, pp. 579–588, 2021

    2021

  3. [11]

    Elevation mapping for locomotion and navigation using gpu,

    T. Miki, L. Wellhausen, R. Grandia, F. Jenelten, T. Homberger, and M. Hutter, “Elevation mapping for locomotion and navigation using gpu,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2022, pp. 2273–2280

    2022

  4. [12]

    Robot-centric elevation mapping with uncertainty estimates,

    P. Fankhauser, M. Bloesch, C. Gehring, M. Hutter, and R. Siegwart, “Robot-centric elevation mapping with uncertainty estimates,” in Mobile Service Robotics. World Scientific, 2014, pp. 433–440

    2014

  5. [13]

    Occupancy-elevation grid: an alternative approach for robotic mapping and navigation,

    A. Souza and L. M. Gonçalves, “Occupancy-elevation grid: an alternative approach for robotic mapping and navigation,” Robotica, vol. 34, no. 11, pp. 2592–2609, 2016

    2016

  6. [14]

    Omni-perception: Omnidirectional collision avoidance for legged locomotion in dynamic environments,

    Z. Wang, T. Ma, Y . Jia, X. Yang, J. Zhou, W. Ouyang, Q. Zhang, and J. Liang, “Omni-perception: Omnidirectional collision avoidance for legged locomotion in dynamic environments,” arXiv preprint arXiv:2505.19214, 2025

    2025 arXiv

  7. [15]

    Agile but safe: Learning collision-free high-speed legged locomotion,

    T. He, C. Zhang, W. Xiao, G. He, C. Liu, and G. Shi, “Agile but safe: Learning collision-free high-speed legged locomotion,” in Robotics: Science and Systems (RSS) , 2024

    2024

  8. [16]

    Coupling vision and proprioception for navigation of legged robots,

    Z. Fu, A. Kumar, A. Agarwal, H. Qi, J. Malik, and D. Pathak, “Coupling vision and proprioception for navigation of legged robots,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17 273–17 283

    2022

  9. [17]

    Safe navigation strategies for a biped robot walking in a crowd,

    N. Bohórquez, A. Sherikov, D. Dimitrov, and P.-B. Wieber, “Safe navigation strategies for a biped robot walking in a crowd,” in 2016 IEEE-RAS 16th International Conference on Humanoid Robots (Humanoids). IEEE, 2016, pp. 379–386

    2016

  10. [18]

    Adapt- ing robot behavior for human–robot interaction,

    N. Mitsunaga, C. Smith, T. Kanda, H. Ishiguro, and N. Hagita, “Adapt- ing robot behavior for human–robot interaction,” IEEE Transactions on Robotics, vol. 24, no. 4, pp. 911–916, 2008

    2008

  11. [19]

    A bellboy robot: Study of the effects of robot behaviour on user engagement and comfort,

    E. Rodriguez-Lizundia, S. Marcos, E. Zalama, J. Gómez-García- Bermejo, and A. Gordaliza, “A bellboy robot: Study of the effects of robot behaviour on user engagement and comfort,” International Journal of Human-Computer Studies , vol. 82, pp. 83–95, 2015

    2015

  12. [20]

    Learning and comfort in human– robot interaction: A review,

    W. Wang, Y . Chen, R. Li, and Y . Jia, “Learning and comfort in human– robot interaction: A review,” Applied Sciences, vol. 9, no. 23, p. 5152, 2019

    2019

  13. [21]

    Customer comfort during service robot interactions,

    M. Becker, D. Mahr, and G. Odekerken-Schröder, “Customer comfort during service robot interactions,” Service Business, vol. 17, no. 1, pp. 137–165, 2023

    2023

  14. [22]

    Anymal parkour: Learning agile navigation for quadrupedal robots,

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter, “Anymal parkour: Learning agile navigation for quadrupedal robots,” Science Robotics, vol. 9, no. 88, p. eadi7566, 2024

    2024

  15. [23]

    Arm-constrained curriculum learning for loco-manipulation of a wheel-legged robot,

    Z. Wang, Y . Jia, L. Shi, H. Wang, H. Zhao, X. Li, J. Zhou, J. Ma, and G. Zhou, “Arm-constrained curriculum learning for loco-manipulation of a wheel-legged robot,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 10 770– 10 776

    2024

  16. [24]

    Soloparkour: Constrained reinforcement learning for visual locomotion from privileged experience,

    E. Chane-Sane, J. Amigo, T. Flayols, L. Righetti, and N. Mansard, “Soloparkour: Constrained reinforcement learning for visual locomotion from privileged experience,” in Conference on Robot Learning (CoRL) , 2024

    2024

  17. [25]

    Navrl: Learning safe flight in dynamic environments,

    Z. Xu, X. Han, H. Shen, H. Jin, and K. Shimada, “Navrl: Learning safe flight in dynamic environments,” IEEE Robotics and Automation Letters, vol. 10, no. 4, pp. 3668–3675, 2025

    2025

  18. [26]

    Safe learning in robotics: From learning-based control to safe reinforcement learning,

    L. Brunke, M. Greeff, A. W. Hall, Z. Yuan, S. Zhou, J. Panerati, and A. P. Schoellig, “Safe learning in robotics: From learning-based control to safe reinforcement learning,” Annual Review of Control, Robotics, and Autonomous Systems , vol. 5, no. 1, pp. 411–444, 2022

    2022

  19. [27]

    A review of safe reinforcement learning: Methods, theory and applications,

    S. Gu, L. Yang, Y . Du, G. Chen, F. Walter, J. Wang, and A. Knoll, “A review of safe reinforcement learning: Methods, theory and applications,” arXiv preprint arXiv:2205.10330 , 2022

    2022 arXiv

  20. [28]

    Penalized proximal policy optimization for safe reinforcement learning,

    L. Zhang, L. Shen, L. Yang, S. Chen, B. Yuan, X. Wang, and D. Tao, “Penalized proximal policy optimization for safe reinforcement learning,” arXiv preprint arXiv:2205.11814 , 2022

    2022 arXiv

  21. [29]

    Safe reinforcement learning in constrained markov decision processes,

    A. Wachi and Y . Sui, “Safe reinforcement learning in constrained markov decision processes,” in International Conference on Machine Learning. PMLR, 2020, pp. 9797–9806

    2020

  22. [30]

    Clf-cbf based quadratic programs for safe motion control of nonholonomic mobile robots in presence of moving obstacles,

    M. Desai and A. Ghaffari, “Clf-cbf based quadratic programs for safe motion control of nonholonomic mobile robots in presence of moving obstacles,” in 2022 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM) . Ieee, 2022, pp. 16–21

    2022

  23. [31]

    Safe and robust motion planning for dynamic robotics via control barrier functions,

    A. Manjunath and Q. Nguyen, “Safe and robust motion planning for dynamic robotics via control barrier functions,” in 2021 60th IEEE Conference on Decision and Control (CDC) . IEEE, 2021, pp. 2122– 2128

    2021

  24. [32]

    Reinforcement learning for safety-critical control under model uncertainty, using control lyapunov functions and control barrier functions,

    J. Choi, F. Castaneda, C. J. Tomlin, and K. Sreenath, “Reinforcement learning for safety-critical control under model uncertainty, using control lyapunov functions and control barrier functions,” arXiv preprint arXiv:2004.07584, 2020

    2004 arXiv

  25. [33]

    Safe reinforcement learning using robust control barrier functions,

    Y . Emam, G. Notomista, P. Glotfelter, Z. Kira, and M. Egerstedt, “Safe reinforcement learning using robust control barrier functions,” IEEE Robotics and Automation Letters , 2022

    2022

  26. [34]

    Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigation

    A. Agrawal and K. Sreenath, “Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigation.” in Robotics: Science and Systems , vol. 13. Cambridge, MA, USA, 2017, pp. 1–10

    2017

  27. [35]

    Learning and comfort in human–robot interaction: A review,

    W. Wang, Y . Chen, R. Li, and Y . Jia, “Learning and comfort in human–robot interaction: A review,” Applied Sciences, vol. 9, no. 23,

  28. [36]

    Perceived safety in physical human–robot interaction—a survey,

    M. Rubagotti, I. Tusseyeva, S. Baltabayeva, D. Summers, and A. Sandygulova, “Perceived safety in physical human–robot interaction—a survey,” Robotics and Autonomous Systems , vol. 151, p. 104047, 2022. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S09...

    2022

  29. [37]

    Influences on proxemic behaviors in human-robot interaction,

    L. Takayama and C. Pantofaru, “Influences on proxemic behaviors in human-robot interaction,” in 2009 IEEE/RSJ international conference on intelligent robots and systems . IEEE, 2009, pp. 5495–5502

    2009

  30. [38]

    Toward safe close- proximity human-robot interaction with standard industrial robots,

    P. A. Lasota, G. F. Rossano, and J. A. Shah, “Toward safe close- proximity human-robot interaction with standard industrial robots,” in 2014 IEEE International Conference on Automation Science and Engineering (CASE), 2014, pp. 339–344

    2014

  31. [39]

    How social distance shapes human–robot interaction,

    Y . Kim and B. Mutlu, “How social distance shapes human–robot interaction,” International Journal of Human-Computer Studies, vol. 72, no. 12, pp. 783–795, 2014

    2014

  32. [40]

    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 Automation Letters...

    2023

  33. [41]

    Genesis: A universal and generative physics engine for robotics and beyond,

    G. Authors, “Genesis: A universal and generative physics engine for robotics and beyond,” December 2024. [Online]. Available: https://github.com/Genesis-Embodied-AI/Genesis

    2024

  34. [2019]

    Available: https://www.mdpi.com/2076-3417/9/23/5152

    [Online]. Available: https://www.mdpi.com/2076-3417/9/23/5152

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.