Pith. sign in

REVIEW 3 major objections 4 minor 27 references

A vision-trained RL policy with a control-barrier safety filter claims collision-proof flight in clutter at up to 7.5 m/s.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 03:10 UTC pith:YO5EVETK

load-bearing objection Solid empirical integration of RL with Dijkstra shaping and a HOCBF filter, but the safety guarantee rests on a variable-type error—the QP constrains acceleration while the policy outputs thrust/attitude—so the 'provably safe' claim is unsupported as written. the 3 major comments →

arxiv 2602.08653 v2 pith:YO5EVETK submitted 2026-02-09 cs.RO

High-Speed Vision-Based Flight in Clutter with Safety-Shielded Reinforcement Learning

Jiarui Zhang , Chengyong Lei , Chengjiang Dai , Kenghou Hoi , Lijie Wang , Zhichao Han , Fei Gao This is my paper
classification cs.RO
keywords quadrotorreinforcement learningcontrol barrier functionsafety filtercollision avoidancehigh-speed flightsim-to-real transferdepth vision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes an end-to-end vision-based quadrotor controller that merges a reinforcement-learning policy with a real-time safety filter. The filter, built from high-order control barrier functions, projects every raw policy command onto a set the authors call provably safe, giving a formal collision-avoidance guarantee even under distribution shift and disturbances. The policy is trained with physics-informed rewards: Dijkstra-based global geodesic guidance to avoid local minima and a control-barrier reward for safety-aware exploration. Benchmark comparisons report higher success rates and closer speed tracking than a traditional planner (Ego-Planner) and a differentiable-physics learning baseline (DiffPhys), and real-world indoor and forest flights are described at speeds up to 7.5 m/s. The significance, if the guarantee holds, is a middle path between brittle pure-learning agility and latency-heavy modular planning.

Core claim

The paper's central claim is that safety can be enforced at deployment time by solving a small quadratic program that minimally modifies the policy's action so that, for every nearby obstacle, the high-order barrier condition \ddot h + \alpha1 \dot h + \alpha0 h \ge 0 is satisfied, where h(r)=||r||^2 - r_safe^2 and r is the relative position to the obstacle. Under a second-order model of the quadrotor, this condition is designed to make the safe set forward invariant, meaning no trajectory that starts outside the collision radius can enter it. The experiments claim that this shielding, combined with Dijkstra-based reward shaping during training, outperforms both a planning baseline and a dif

What carries the argument

The HOCBF safety filter. For each obstacle, the paper defines h_i(r_t)=||r_t||^2 - r_safe^2, then forms the QP in Eq. (6)-(8): minimize ||a - a_raw||^2 subject to the linear constraint 2 r_t·a \ge -2||v_t||^2 - \alpha1 \dot h_i - \alpha0 h_i, obtained by substituting \dot h_i = 2 r_t·v_t and \ddot h_i = 2||v_t||^2 + 2 r_t·a. The QP is intended to project the policy's raw command onto the provably safe set in real time. The training-side mechanisms — Dijkstra geodesic rewards and CBF-based safety rewards — are meant to make the policy aligned with the filter so that corrections stay small.

Load-bearing premise

The certificate assumes the variable the safety filter corrects — the drone's translational acceleration — is effectively the variable the policy commands; the policy actually sends a thrust magnitude and attitude angles that a low-level controller must track, so the forward-invariance proof does not cover the real actuation chain, especially with 60-80 ms latency and noisy depth input.

What would settle it

Run the trained policy with the HOCBF filter in the paper's simulator, replace the policy output with random actions (or fixed forward thrust), keep the same 60-80 ms latency and depth noise, and check whether any collision occurs. If a collision is observed in a scene where the QP constraints are feasible for the true dynamics, the claimed forward invariance is falsified; likewise, comparing commanded acceleration in the QP with IMU-measured acceleration during the real flights would show whether the constraint is actually enforced.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • End-to-end learned policies could be deployed in safety-critical flight without a separate receding-horizon planner, cutting latency while keeping a mathematical safety net.
  • The success-rate numbers imply that hybrid shielding extends the usable speed envelope of vision-based RL past what soft constraints alone achieve at 7-9 m/s.
  • Because the filter runs at 100 Hz and tolerates emulated 60-80 ms delay, the approach is compatible with the latency budgets of real onboard computers and flight controllers.
  • If the advertised guarantee transfers, the same QP-shielding recipe can be reused for other perception-action loops where the protected quantity (e.g., distance to obstacle) is computable online.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's formal claim is only as strong as the link between the QP's decision variable (translational acceleration) and the actual action (mass-normalized thrust plus attitude setpoints). A stricter certificate would require a control-barrier condition on the attitude-controlled dynamics or an outer loop designed so the attitude controller translates the filter's corrections into true accelerat
  • A natural test of the filter's independent contribution would be to run it on a random or fixed policy in simulation; the current ablations only show the full system against variants without the filter, leaving the filter's safety gain entangled with the learned policy.
  • The reported 7.5 m/s forest results suggest the distribution-shift protection might extend to novel obstacle geometries beyond the training set, but dynamic obstacles and fine-scale obstacles, which the paper lists as future work, would stress whether the point-cloud-based h_i construction is sufficient.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes a hybrid framework for high-speed vision-based quadrotor flight in cluttered environments. During training, a PPO policy is shaped by a Dijkstra-geodesic navigation reward and a CBF-based safety reward. At deployment, a high-order control barrier function (HOCBF) quadratic program is used as a safety filter that projects the policy's raw 4D action (normalized thrust and attitude angles) onto a supposedly 'provably safe set'. The authors report benchmark comparisons against Ego-Planner and DiffPhys, plus indoor and outdoor real-world flights at speeds up to 7.5 m/s, and claim that the HOCBF filter 'guarantees collision avoidance under unforeseen disturbances and distribution shifts'.

Significance. If the formal safety guarantee were valid, this would be a valuable contribution: an end-to-end vision-based controller with a real-time, provably safe correcting filter, demonstrated at high speed in real environments. The training-time Dijkstra-guided reward is an interesting idea for avoiding local minima. However, the central formal claim is not supported by the presented derivation, and the empirical validation is insufficiently specified to substantiate the performance claims. The paper is therefore not publishable in its current form.

major comments (3)
  1. [§III-D, Eq. (8)] The HOCBF constraint in Eq. (7)–(8) treats a_t as a 3D translational acceleration vector: it substitutes ddot h_i = 2||v_t||^2 + 2 r_tᵀ a_t and then writes the linear constraint 2 r_tᵀ a_t ≥ ... . However, the policy action a_t ∈ R^4 defined in Eq. (2) is [f_n, φ_ref, θ_ref, ψ_ref]ᵀ, i.e., mass-normalized thrust and attitude angle references. The dot product r_tᵀ a_t is dimensionally inconsistent (3-vector times 4-vector), and the actual translational acceleration is determined by the attitude dynamics and the low-level controller, not by a_t directly. The QP therefore does not constrain the commanded quantity, and the claimed forward invariance of the safe set is not established. This is a fundamental flaw in the safety certificate that is the paper's core contribution.
  2. [§III-D, §III-E] Even setting aside the dimension mismatch, the safety filter is derived for a perfect, zero-latency state feedback system with known obstacle geometry. The deployment pipeline, as described in Sec. III-E and the real-world experiments, uses depth images reconstructed into point clouds, subject to sensor noise, pixel dropout, and 60–80 ms latency. None of these uncertainties or delays appear in the HOCBF model. Thus the statement in the Abstract and Sec. I that the filter guarantees collision avoidance 'under unforeseen disturbances and distribution shifts' is not supported by the analysis. At best, the filter enforces a kinematic condition on an idealized model with perfect perception; the paper should clearly state these limitations.
  3. [Tables I and II, §III-A3] The empirical claims of superior performance are not adequately supported. Table II reports success rates based on only 10 trials per configuration, yet no confidence intervals or per-trial results are given; differences of 10–20 percentage points are within sampling noise. Furthermore, the reward weights α_i, the Dijkstra scaling λ and clipping bound C, the CBF safety margin d_safe and coefficient γ, the HOCBF coefficients α_0 and α_1, and the deployment safety radius r_safe are never reported. These parameters are all empirically tuned, so the results cannot be reproduced or independently verified. This is particularly important because the safety filter's behavior depends directly on r_safe and α_0, α_1.
minor comments (4)
  1. [Abstract and throughout] Typographical and stylistic issues: 'UA Vs' should be 'UAVs'; 'Diffphys' and 'DiffPhys' are used inconsistently; 'Ego-PlannerV2' should be 'Ego-Planner V2'; reference [26] lists only 'Xiao and Belta' without initials.
  2. [§III-C, Eq. (4)] The derivation of r_safety in Eq. (5) uses h(x) = d(x) − d_safe, but the clip lower bound δ_min is set to −2.0. The reason for this specific value and how it interacts with the CBF condition is not discussed. Please clarify the units and the effect on learning.
  3. [§IV-A3] The benchmark comparison with Ego-Planner reports 0% success at ≥5 m/s. Given that Ego-Planner is a local planner designed for slower flight, this baseline choice should be justified. The comparison would benefit from reporting the computational latency and planning frequency of each baseline.
  4. [§IV-B] The real-world experiments are described qualitatively. Provide quantitative results: number of flights, success rates, flight time, and a measure of closeness to obstacles. The outdoor forest flight claims an average velocity of 7 m/s over 35 m; specify the peak speed and the obstacle density.

Circularity Check

0 steps flagged

No significant circularity; the main weakness is a correctness gap in the HOCBF filter, not a circular derivation.

full rationale

I examined the paper for the seven enumerated circularity patterns. No fitted quantity is disguised as a prediction: the reward weights, CBF coefficients, and safety radii are described as empirically tuned (Eq. 3, Eq. 7) and are not presented as derived predictions. The HOCBF safety filter (Eq. 6-8) is an application of external, machine-checkable control barrier function theory [23]-[26]; the derivation of the constraint from h_i = ||r_t||^2 - r_safe^2 and its derivatives is a standard calculation under the double-integrator assumption, not an input-equivalent redefinition. The benchmarks compare against external baselines (Ego-Planner, DiffPhys) and are not statistically forced by the method's own parameters. Self-citations ([6], [16], [17], [20]) appear only as contextual examples of learning-based or modular navigation, not as load-bearing justifications for the central claim. There is no uniqueness theorem imported from the authors' prior work and no ansatz smuggled in via self-citation. I note, however, a serious correctness issue that is distinct from circularity: Eq. (2) defines the action a_t as a 4D vector [f_n, phi_ref, theta_ref, psi_ref]^T, while Eq. (8) uses a_t in the 3D dot product r_t^T a_t and treats it as translational acceleration. The dimensions are inconsistent, and the actual attitude-controlled dynamics with 60-80 ms latency are not the assumed double-integrator dynamics. This means the claimed forward-invariance guarantee is not established by the presented derivation, but this is an invalid application of external theory, not a reduction of the result to its own inputs. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The central safety guarantee rests on an unstated double-integrator approximation and on perception reliability; most tuning parameters are unreported, pushing the implementation burden onto the reader.

free parameters (7)
  • Reward weights alpha_1..alpha_7 = not reported
    Empirically tuned coefficients in Eq. (3); values are not given.
  • Dijkstra reward scaling lambda and clipping bound C = not reported
    Parameters of Eq. (4); values not reported.
  • CBF safety reward margin d_safe and coefficient gamma = not reported
    Parameters of the barrier function and Eq. (5); values not reported.
  • Safety reward clipping threshold delta_min = -2.0
    Stated in Sec. III-C; a hand-set bound.
  • HOCBF coefficients alpha_0 and alpha_1 = not reported
    Adjustable coefficients in Eq. (7); values not reported.
  • Deployment safety radius r_safe = not reported
    Safety radius in the HOCBF barrier function; value not reported.
  • Domain randomization ranges (delay 60-80 ms, ±10% action noise, dropout rates) = reported ranges
    Hand-chosen sim-to-real randomization settings in Sec. III-E; ranges given but no sensitivity analysis.
axioms (4)
  • domain assumption The policy action a_t is the translational acceleration of the quadrotor in the HOCBF constraint.
    Eq. (8) uses a_t as acceleration in \ddot h_i, while Eq. (2) defines a_t as thrust/attitude references; this assumption is unstated and inconsistent with the action space.
  • domain assumption The nearest obstacle point r_t from the local point cloud is accurate and persistent at 100 Hz.
    The HOCBF guarantee depends on a fixed relative position to the nearest obstacle point; noisy or switching detections would break the forward-invariance argument.
  • domain assumption A precomputed Dijkstra cost map per scene is available during training.
    The navigation reward in Algorithm 1 requires a per-scene geodesic grid, which limits applicability to known training environments.
  • domain assumption ESDF gradients used in the safety reward are reliable approximations of distance derivatives.
    r_safety in Eq. (5) uses \nabla d(x) · v_t from a constructed ESDF; inaccuracies in the ESDF would feed incorrect safety shaping.

pith-pipeline@v1.3.0-alltime-deepseek · 10805 in / 10292 out tokens · 110905 ms · 2026-08-03T03:10:29.194346+00:00 · methodology

0 comments
read the original abstract

Quadrotor unmanned aerial vehicles (UAVs) are increasingly deployed in complex missions that demand reliable autonomous navigation and robust obstacle avoidance. However, traditional modular pipelines often incur cumulative latency, whereas purely reinforcement learning (RL) approaches typically provide limited formal safety guarantees. To bridge this gap, we propose an end-to-end RL framework augmented with model-based safety mechanisms. We incorporate physical priors in both training and deployment. During training, we design a physics-informed reward structure that provides global navigational guidance. During deployment, we integrate a real-time safety filter that projects the policy outputs onto a provably safe set to enforce strict collision-avoidance constraints. This hybrid architecture reconciles high-speed flight with robust safety assurances. Benchmark evaluations demonstrate that our method outperforms both traditional planners and recent end-to-end obstacle avoidance approaches based on differentiable physics. Extensive experiments demonstrate strong generalization, enabling reliable high-speed navigation in dense clutter and challenging outdoor forest environments at velocities up to 7.5 m/s}.

Figures

Figures reproduced from arXiv: 2602.08653 by Chengjiang Dai, Chengyong Lei, Fei Gao, Jiarui Zhang, Kenghou Hoi, Lijie Wang, Zhichao Han.

Figure 1
Figure 1. Figure 1: Snapshot of real-world indoor flight experiments. Flight trajectories recorded via long-exposure photography demonstrate the UAV’s agile maneuvering capabilities during high-speed traversal between trees and artificial obstacles. tional costs to scale poorly with environmental complexity and limiting real-time performance at high flight speeds. In contrast, learning-based navigation has emerged as an attra… view at source ↗
Figure 2
Figure 2. Figure 2: Network architecture and control pipeline. An asymmetric actor-critic policy fuses a depth image (CNN) and proprioceptive states via a GRU. The actor outputs attitude references and normalized thrust, which are tracked by a PX4 controller. During training, domain randomization (dropout/noise/delay) improves robustness; during deployment, a real-time HOCBF filter refines the raw command to enforce safety co… view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of depth observations. (a) Raw depth image [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training performance and environment. (a) Comparative training curves displaying success rates over iterations for different reward configurations. (b) Visualization of a representative training scenario populated with dense obstacles. complex non-convex topologies. We execute 50 independent trials per scene, with the results summarized in Tab. I. Overall, the results strongly validate our hierarchical des… view at source ↗
Figure 5
Figure 5. Figure 5: Benchmark comparisons against state-of-the-art methods. (a) Average realized velocity of Ego-Planner [14], DiffPhys [11], and our method across varying target speeds and environmental configurations. (b) Visualization of the resulting flight trajectories generated by each method in geometric clutter (left) and Perlin noise (right) scenarios. to stabilize visual inputs, thereby severely limiting the realize… view at source ↗
Figure 6
Figure 6. Figure 6: Snapshots of indoor real-world experiments at varying target [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of the outdoor flight experiment. (a) Global point cloud map overlaid with the executed trajectory and the corresponding realized speed profile. (b) Egocentric observations, including RGB and aligned depth images (Intel RealSense D435i), augmented with real-time attitude and speedometer displays. (c) Top￾down view of the environment. (d) Side-view composite image illus￾trating the sequential … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 5 linked inside Pith

  1. [1]

    FUEL: Fast uav exploration using incremental frontier structure and hierarchical planning,

    B. Zhou, Y . Zhang, X. Chen, and F. Gao, “FUEL: Fast uav exploration using incremental frontier structure and hierarchical planning,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 779–786, 2021

    2021

  2. [2]

    A vision-based autonomous uav inspection framework for unknown tunnel construction sites with dynamic obstacles,

    Z. Xu, B. Chen, X. Zhan, Y . Xiu, C. Suzuki, and K. Shimada, “A vision-based autonomous uav inspection framework for unknown tunnel construction sites with dynamic obstacles,”IEEE Robotics and Automation Letters, vol. 8, no. 8, pp. 4983–4990, 2023

    2023

  3. [3]

    Robust and effi- cient quadrotor trajectory generation for fast autonomous flight,

    B. Zhou, F. Gao, L. Wang, C. Liu, and S. Shen, “Robust and effi- cient quadrotor trajectory generation for fast autonomous flight,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 3529–3536, 2019

    2019

  4. [4]

    Teach-repeat- replan: A complete and robust system for aggressive flight in complex environments,

    F. Gao, L. Wang, B. Zhou, X. Zhou, J. Pan, and S. Shen, “Teach-repeat- replan: A complete and robust system for aggressive flight in complex environments,”IEEE Transactions on Robotics, vol. 36, no. 5, pp. 1526– 1545, 2020

    2020

  5. [5]

    Swarm of micro flying robots in the wild,

    X. Zhou, X. Wen, Z. Wang, Y . Gao, H. Li, Q. Wang, T. Yang, H. Lu, Y . Cao, C. Xuet al., “Swarm of micro flying robots in the wild,”Science Robotics, vol. 7, no. 66, p. eabm5954, 2022

    2022

  6. [6]

    Whole-body control through narrow gaps from pixels to action,

    T. Wu, Y . Chen, T. Chen, G. Zhao, and F. Gao, “Whole-body control through narrow gaps from pixels to action,”arXiv preprint arXiv:2409.00895, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Badgr: An autonomous self- supervised learning-based navigation system,

    G. Kahn, P. Abbeel, and S. Levine, “Badgr: An autonomous self- supervised learning-based navigation system,”IEEE Robotics and Au- tomation Letters, vol. 6, no. 2, pp. 1312–1319, 2021

    2021

  8. [8]

    Motion primitives- based navigation planning using deep collision prediction,

    H. Nguyen, S. H. Fyhn, P. De Petris, and K. Alexis, “Motion primitives- based navigation planning using deep collision prediction,” in2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 9660–9667

    2022

  9. [9]

    Learning perception-aware agile flight in cluttered environments,

    Y . Song, K. Shi, R. Penicka, and D. Scaramuzza, “Learning perception-aware agile flight in cluttered environments,”arXiv preprint arXiv:2210.01841, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    Mavrl: Learn to fly in cluttered environments with varying speed,

    H. Yu, C. De Wagter, and G. C. E. de Croon, “Mavrl: Learn to fly in cluttered environments with varying speed,”IEEE Robotics and Automation Letters, 2024

    2024

  11. [11]

    Learning vision-based agile flight via differentiable physics,

    Y . Zhang, Y . Hu, Y . Song, D. Zou, and W. Lin, “Learning vision-based agile flight via differentiable physics,”Nature Machine Intelligence, pp. 1–13, 2025

    2025

  12. [12]

    Learning high-speed flight in the wild,

    A. Loquercio, E. Kaufmann, R. Ranftl, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Learning high-speed flight in the wild,”Science Robotics, vol. 6, no. 59, p. eabg5810, 2021

    2021

  13. [13]

    Vision-controlled micro flying robots: from system design to autonomous navigation and mapping in gps-denied environ- ments,

    D. Scaramuzza, M. C. Achtelik, L. Doitsidis, F. Friedrich, E. Kos- matopoulos, A. Martinelli, M. W. Achtelik, M. Chli, S. Chatzichristofis, L. Kneipet al., “Vision-controlled micro flying robots: from system design to autonomous navigation and mapping in gps-denied environ- ments,”IEEE Robotics & Automation Magazine, vol. 21, no. 3, pp. 26–40, 2014

    2014

  14. [14]

    Ego-planner: An esdf- free gradient-based local planner for quadrotors,

    X. Zhou, Z. Wang, H. Ye, C. Xu, and F. Gao, “Ego-planner: An esdf- free gradient-based local planner for quadrotors,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 478–485, 2020

    2020

  15. [15]

    Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex environments,

    S. Liu, M. Watterson, K. Mohta, K. Sun, S. Bhattacharya, C. J. Taylor, and V . Kumar, “Planning dynamically feasible trajectories for quadrotors using safe flight corridors in 3-d complex environments,”IEEE Robotics and Automation Letters, vol. 2, no. 3, pp. 1688–1695, 2017

    2017

  16. [16]

    Reactive aerobatic flight via reinforcement learning,

    Z. Han, X. Huang, Z. Xu, J. Zhang, Y . Wu, M. Wang, T. Wu, and F. Gao, “Reactive aerobatic flight via reinforcement learning,”arXiv preprint arXiv:2505.24396, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Flying on point clouds with reinforcement learning,

    G. Xu, T. Wu, Z. Wang, Q. Wang, and F. Gao, “Flying on point clouds with reinforcement learning,”arXiv preprint arXiv:2503.00496, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    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, 2025

    2025

  19. [19]

    Learning speed adaptation for flight in clutter,

    G. Zhao, T. Wu, Y . Chen, and F. Gao, “Learning speed adaptation for flight in clutter,”IEEE Robotics and Automation Letters, vol. 9, no. 8, pp. 7222–7229, 2024

    2024

  20. [20]

    Dynamically feasible trajectory gen- eration with optimization-embedded networks for autonomous flight,

    Z. Han, L. Xu, L. Pei, and F. Gao, “Dynamically feasible trajectory gen- eration with optimization-embedded networks for autonomous flight,” IEEE Robotics and Automation Letters, 2025

    2025

  21. [22]

    Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,

    Y . Song, A. Romero, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,”Science Robotics, vol. 8, no. 82, p. eadg1462, 2023

    2023

  22. [23]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,”IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2016

    2016

  23. [24]

    Control barrier functions: Theory and applications,

    A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control barrier functions: Theory and applications,” in2019 18th European control conference (ECC). Ieee, 2019, pp. 3420–3431

    2019

  24. [25]

    End-to-end safe reinforcement learning through barrier functions for safety-critical continuous control,

    R. Cheng, G. Orosz, R. M. Murray, and J. W. Burdick, “End-to-end safe reinforcement learning through barrier functions for safety-critical continuous control,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 33, no. 01, 2019, pp. 3387–3395

    2019

  25. [26]

    High-order control barrier functions,

    Xiao and Belta, “High-order control barrier functions,”IEEE Transac- tions on Automatic Control, vol. 67, no. 7, pp. 3655–3662, 2021

    2021

  26. [27]

    Navier-stokes, fluid dynamics, and image and video inpainting,

    M. Bertalmio, A. L. Bertozzi, and G. Sapiro, “Navier-stokes, fluid dynamics, and image and video inpainting,” inProceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, vol. 1. IEEE, 2001, pp. I–I

    2001

  27. [2017]

    Available: http://arxiv.org/abs/1707.06347

    [Online]. Available: http://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv