Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Uncertainty-Aware Safety-Critical Decision and Control for Autonomous Vehicles at Unsignalized Intersections

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The USDC framework cuts autonomous-vehicle collision rates at unsignalized intersections to 3.7 percent while keeping average speed near 7.8 m/s.

desk verdict A plausible integration of ensemble distributional RL with uncertainty-tightened HOCBF, with a real simulation win but an underspecified safety filter that is not a certified guarantee. read the letter →

arxiv 2505.19939 v2 pith:UIFCLG7F submitted 2025-05-26 cs.RO

Ran Yu , Zhuoren Li , Lu Xiong , Wei Han , Bo Leng This is my paper · ORCID
classification cs.RO
keywords autonomousdrivingunsignalizedintersectionreinforcementlearningcontrolbarrierfunctionuncertaintyquantificationdistributionalconditionalvalue-at-risksafetyfilter
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 proposes the USDC framework, which wraps a risk-averse reinforcement-learning policy in a safety filter that adapts to how uncertain the policy is about its decisions. The filter is a high-order control barrier function whose constraints tighten when an ensemble of distributional critics signals high epistemic or aleatoric uncertainty. The same ensemble is used to switch between the filtered and unfiltered actions based on the percentile of current joint uncertainty. In simulated unsignalized intersections, USDC reduces the average collision rate to 3.7 percent, down from 14.7 percent for the DSAC baseline, while keeping average episode velocity at 7.77 m/s versus 7.60 m/s for DSAC.

What carries the argument

The load-bearing object is the joint uncertainty $\sigma_{JU}$, computed from a deep ensemble of distributional critics, each built as a quantile network plus a fixed randomized prior network and trained with bootstrap masks. $\sigma_{JU}$ is a softmax-weighted combination of epistemic uncertainty (disagreement among critics' CVaR estimates) and aleatoric uncertainty (per-critic return variance). It plays two roles: it tightens the HOCBF safety constraint through a positively correlated term, so higher uncertainty means more conservative corrections, and it sets the probability threshold for switching between the RL action and the filtered action. The filter itself is a Truncated Taylor CBF of relative degree 2, based only on relative distance to obstacle envelopes and road boundaries, solved as a quadratic optimization.

What would settle it

Run the trained USDC agent in the same simulator but apply a sudden 3 m/s2 deceleration to an oncoming vehicle during the filter's control interval, at the moment the ego vehicle passes the commitment point; if the relative-distance constraint is violated and a collision follows, the stated safety guarantee is not robust to obstacle motion.

Watch

Extended reading notes

Core claim

The central claim is that estimating both epistemic and aleatoric uncertainty, then feeding them into both the RL objective and the safety filter, resolves the usual safety-efficiency trade-off in intersection driving. The paper shows that distributional critics can compute a CVaR-based value for both the RL action and the filtered action, and that selecting the RL action only when a majority of critics judge it no worse, and only when current uncertainty is below its historical percentile, yields lower collision rates than pure RL without sacrificing speed. Across left-turn, straight-through, and right-turn tasks, USDC's mean collision rate is 3.7 percent versus 14.7 percent for DSAC, and its mean episode velocity is 7.77 m/s versus 7.60 m/s.

Load-bearing premise

The safety filter's guarantee depends on obstacle points being treated as fixed while deriving the distance constraint derivatives, so a surrounding vehicle that accelerates or turns after the ego vehicle commits can violate the constraint before the next control step; the paper itself notes the HOCBF only guarantees safety for a single step and ignores obstacle-state changes.

Editorial extensions

If this is right

  • Collision rate across left-turn, straight, and right-turn tasks falls to 3.7 percent on average, an 11-point reduction over DSAC.
  • Average episode velocity stays at 7.77 m/s, within 3 percent of DSAC, indicating the safety gain does not come from conservative crawling.
  • The always-on CBF baseline achieves 2.8 percent collisions but a 27.8 percent frozen rate and lower reward, implying the switching mechanism is responsible for efficiency, not the filter alone.
  • The HOCBF needs only relative distances, not full surrounding-vehicle state, making the approach transferable to other traffic geometries.
  • Under low uncertainty the RL action is chosen when its CVaR is judged no worse, while under high uncertainty the safer filtered action is favored.

Reading between the lines

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

  • Because the paper derives HOCBF constraint derivatives without obstacle velocity and acceleration terms, the most informative next experiment would vary obstacle accelerations during filter interventions; if collisions reappear, the single-step guarantee is the actual limit.
  • A testable extension is to feed the ensemble's disagreement into the control frequency: at high $\sigma_{JU}$, lowering the policy execution frequency would shorten the horizon the filter must compensate.
  • The percentile-based switching rule could be generalized to continuous action mixing rather than hard switching, potentially smoothing interventions at the boundary.
  • The same joint-uncertainty-driven filter could be applied to other relative-degree-2 constraints, such as lane-keeping or merge spacing, with little modification.
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

2 major / 5 minor

Summary. The paper proposes the USDC framework for autonomous driving at unsignalized intersections. It combines a risk-averse ensemble distributional RL policy (using CVaR, randomized prior functions, and bootstrapped critics) with a high-order control barrier function (HOCBF) safety filter whose constraints are tightened by a joint uncertainty measure, plus an ensemble-critic-based switching rule that selects between the RL and CBF actions. The framework is evaluated in a Highway-Env intersection simulator on left-turn, straight, and right-turn tasks against DSAC, Pure-DSAC, Dir-CBF, and USDC without the CBF. The reported mean collision rates are 3.7% for USDC versus 14.7% for DSAC and 2.8% for Dir-CBF, with average episode velocities of 7.77 m/s, 7.60 m/s, and 6.79 m/s respectively, supporting the paper's claim of improved safety with maintained traffic efficiency relative to the RL baselines.

Significance. The problem is relevant and the empirical setup is nontrivial: multiple driving tasks, several baselines, 200 test episodes per task, and repeated training runs with different seeds. The explicit limitations stated in the conclusion—that the HOCBF only guarantees safety for a single step and does not account for changes in obstacle states—are honest but they directly undercut the safety-filter claim. If the filter derivation and the uncertainty-tightening function were completed and validated, the framework could be a useful combination of distributional RL and CBF-based safety filtering. As written, the reported collision reduction cannot yet be attributed to the stated safety mechanism, and the central component W in Eq. (30) is not reproducible.

major comments (2)
  1. [§III-B.1, Eq. (24)–(26)] The derivative computation for h_veh treats obstacle points as fixed. With h_veh = (x_k - x_obs,k)^2 + (y_k - y_obs,k)^2 - (r_obs + r_ego)^2, the true first derivative contains the obstacle velocity terms -2(x_k - x_obs,k)v_obs,x - 2(y_k - y_obs,k)v_obs,y, and the second derivative contains obstacle acceleration terms. At the test frequency fπ = 10 Hz, an SV moving at 6–10 m/s changes the relative position by 0.6–1.0 m per policy step, the same order as the envelope radius and the slack term α(h). The constraint in Eq. (27) is therefore optimistic when an obstacle is approaching, which is exactly the situation the filter is supposed to handle. The conclusion states that the HOCBF 'only guarantees safety for a single step and does not account for the effects of changes in obstacle states.' Since the central claim is that USDC improves safety, the low CR in Table I cannot be credited to the safety filter without either incorporating obstacle dynamics into the TTCBF or re-framing the filter as a heuristic and providing separate evidence for the safety claim.
  2. [§III-B.2, Eq. (30)] The uncertainty-tightened constraint is not reproducible as written. The text replaces ΓΔt^{r+1} - w(x_k,u_k) with W(x_k,u_k,σ_JU), described only as 'a positively correlated function of σ_JU.' No functional form, parameters, or tuning procedure is given, and the optimization in Eq. (28) is not shown with W. Because this substitution is the mechanism that makes the filter more conservative under high uncertainty and less conservative under low uncertainty, the reported safety/efficiency balance depends on an unspecified design choice. Please provide the explicit form of W, its units, and a sensitivity study over its parameters.
minor comments (5)
  1. [§II-A, before Eq. (4)] The sentence 'We then tarin the critic' contains a typo; it should read 'We then train the critic.'
  2. [Definition 4] 'Given a serious sets Ci' should read 'Given a series of sets Ci.'
  3. [Eq. (22)] In the definition of U(x), the symbols µ and η are described as the mean and variance of σ_AU or σ_EU, but η appears to be a scale parameter rather than a variance; please clarify the notation.
  4. [Eq. (31)] The symbol P is used for the proportion of ensemble critics that prefer u_RL over u_CBF, which can be confused with transition probabilities; using a different notation, such as an explicit fraction or count, would improve clarity.
  5. [§III-B.2, Eq. (29)] The notation φ(x,u) ∈ X × U ⊂ R^n is dimensionally inconsistent; φ is a function from X × U to R^n, not an element of X × U.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the framework is empirically benchmarked against external baselines, and the only self-citation is for a network component and metric definitions, not for a load-bearing proof step.

full rationale

The central claim, that USDC reduces collision rate while maintaining average episode speed relative to DSAC, Pure-DSAC, Dir-CBF, and USDC without the filter, is supported by direct simulation comparisons in Table I, not by a derivation whose conclusion is reused as an input. The HOCBF safety filter follows the standard discrete-time CBF/TTCBF construction: the candidate barrier is defined in Eq. (24), its derivatives are given in Eq. (26), and the filter optimization is stated in Eq. (28). The derivatives in Eq. (26) treat obstacle points as fixed while surrounding vehicles move, a correctness limitation that the authors explicitly concede: 'the HOCBF in this paper only guarantees safety for a single step and does not account for the effects of changes in obstacle states.' That limitation is serious but is not circularity, because the safety claim is not being proven from an assumed version of itself. The uncertainty-tightened constraint in Eq. (30) replaces a remainder term with an under-specified 'positively correlated function' W(x_k,u_k,sigma_JU); this is a reproducibility and rigor concern, not a fitted parameter disguised as a prediction. The switching rule in Eq. (31) is a heuristic selection between the RL action and the filtered action based on ensemble-critic CVaR estimates; the statement that the RL policy 'performs no worse than the HOCBF under low uncertainty' is an interpretation of that selection rule rather than an independently predicted outcome, and the reported collision and speed numbers in Table I are measured outcomes, not consequences of Eq. (31) alone. The only self-citation is [40], used for the two-hop attention encoder in Section IV-C ('Based on our previous work [40]') and for the Table I metric definitions ('The definitions are similar to those in [40]'); this is a component-level dependency and does not carry the paper's safety or efficiency conclusions. No step in the derivation chain reduces by construction to a self-cited theorem or to a fitted parameter, so the appropriate circularity score is low.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central safety mechanism depends on several unspecified hyperparameters and on an ad hoc mapping from uncertainty to constraint tightening. The HOCBF itself is standard, but the uncertainty-embedded version is not derived.

free parameters (7)
  • rho (prior factor) = not specified
    Controls the weight of the fixed prior network in each ensemble critic, Eq. (16). The value affects epistemic uncertainty estimates and the resulting policy.
  • beta (CVaR quantile) = not specified
    Sets the tail risk level in the risk-sensitive actor objective, Eq. (19). Conservative behavior depends on this choice.
  • Gamma (TTCBF remainder bound) = not specified
    Must dominate the Taylor remainder in Eq. (25) to make the truncated barrier valid. The paper does not report how it is chosen.
  • lambda or alpha_i (class K functions) = not specified
    The discrete-time HOCBF condition uses linear class K functions with parameters that are not reported.
  • W(x,u,sigma_JU) (uncertainty constraint) = not specified
    The central mechanism that tightens the safety constraint under high uncertainty is described only as a positively correlated function, with no explicit formula.
  • N (ensemble size) = not specified
    The number of ensemble critics determines uncertainty estimates and the switching probability in Eq. (31), but is not stated.
  • Reward weights Q, R_a, R_Delta = Q=diag(400,...,0.5), R_a=diag(0.05,0.02), R_Delta=[0.2,0.3]
    Hand-tuned to shape the RL policy; they influence which behaviors are learned, though they are not the safety filter itself.
assumptions (5)
  • ad hoc to paper The joint uncertainty sigma_JU is a sufficient proxy for the unknown disturbance w in the robust HOCBF condition.
    Eq. (22) combines aleatoric and epistemic uncertainties into a scalar, and after Eq. (30) the robust term is replaced by a positively correlated function of sigma_JU without derivation or validation.
  • domain assumption Obstacle positions are treated as fixed when computing relative-distance derivatives.
    Eq. (26) computes dot h and double dot h without obstacle velocity or acceleration terms; the conclusion acknowledges obstacle state changes are not accounted for.
  • domain assumption The kinematic bicycle model (Eq. 23) describes the ego vehicle motion accurately.
    All HOCBF constraints and the simulation controller rely on this approximate model.
  • standard math The TTCBF Taylor remainder bound Gamma is valid and finite.
    Definition 5 requires Gamma such that Gamma/(r+1)! Delta t^{r+1} >= |R_{r+1}|; the paper does not show how to compute Gamma for the nonlinear system.
  • domain assumption The Highway-Env traffic model with IDM is an adequate evaluation environment.
    All performance claims are based on this simulator, not on real-world or more realistic sensor data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Safety-Critical Decision and Control for Autonomous Vehicles at Unsignalized Intersections." pith.science (2026). https://pith.science/paper/UIFCLG7F

@misc{pith2026250519939,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Safety-Critical Decision and Control for Autonomous Vehicles at Unsignalized Intersections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UIFCLG7F}},
  note         = {Machine review of arXiv:2505.19939}
}
read the original abstract

Reinforcement learning (RL) has demonstrated potential in autonomous driving (AD) decision tasks. However, applying RL to urban AD, particularly in intersection scenarios, still faces significant challenges. The lack of safety constraints makes RL vulnerable to risks. Additionally, cognitive limitations and environmental randomness can lead to unreliable decisions in safety-critical scenarios. Therefore, it is essential to quantify confidence in RL decisions to improve safety. This paper proposes an Uncertainty-aware Safety-Critical Decision and Control (USDC) framework, which generates a risk-averse policy by constructing a risk-aware ensemble distributional RL, while estimating uncertainty to quantify the policy's reliability. Subsequently, a high-order control barrier function (HOCBF) is employed as a safety filter to minimize intervention policy while dynamically enhancing constraints based on uncertainty. The ensemble critics evaluate both HOCBF and RL policies, embedding uncertainty to achieve dynamic switching between safe and flexible strategies, thereby balancing safety and efficiency. Simulation tests on unsignalized intersections in multiple tasks indicate that USDC can improve safety while maintaining traffic efficiency compared to baselines.

Figures

Figures reproduced from arXiv: 2505.19939 by the authors.

Figure 1
Figure 1. Diagram of our framework. where ∆h(xk,uk) = h(xk+1) − h(xk). Note that linear class K function is commonly used in the discrete domain, i.e., α(h(x)) = λh(x), λ ∈ (0, 1]. Definition 3 (Relative degree): A continuously differen￾tiable function h : X → R is said to have relative degree r ∈ N with respect to system (8), if ∀x ∈ X , LgL i fh(x) = 0, ∀i ∈ [0, 1, ..., r − 2], and LgL r−1 f h(x) ̸= 0. The standard CBF assu… view at source ↗
Figure 2
Figure 2. Diagram of relative distance. At the k-th time step, the envelopes of the ego vehicle (EV) and surrounding vehicles (SV) are represented by three circles, denoted as (x ego i|k , y ego i|k ) and (x obs i|k , yobs i|k ), i ∈ {1, 2, 3}. Then, we select N closest points to the EV, (x obs j|k , yobs j|k ), j ∈ {1, ..., N}, and M constraints related to the road boundaries. The construction of distance constraints are for… view at source ↗
Figure 3
Figure 3. Details of designed network. The components within the dashed [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

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. An interactive enhanced driving dataset for autonomous driving

    cs.CV 2026-02 conditional novelty 5.0 of 10

    A fused, interaction-labeled dataset of 7.31M driving segments with synthetic BEV videos and VQA pairs for training/evaluating driving VLMs.

Reference graph

Works this paper leans on

40 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    A survey on an emerging safety challenge for autonomous vehicles: Safety of the intended functionality,

    H. Wang, W. Shao, C. Sun, K. Yang, D. Cao, and J. Li, “A survey on an emerging safety challenge for autonomous vehicles: Safety of the intended functionality,” Engineering, vol. 33, pp. 17–34, 2024

    work page 2024

  2. [2]

    A matched case-control analysis of au- tonomous vs human-driven vehicle accidents,

    M. Abdel-Aty and S. Ding, “A matched case-control analysis of au- tonomous vs human-driven vehicle accidents,” Nat. Commun., vol. 15, no. 1, p. 4931, 2024

    work page 2024

  3. [3]

    Decision-making framework for autonomous driving at road intersections: Safeguarding against collision, overly conservative behavior, and violation vehicles,

    S. Noh, “Decision-making framework for autonomous driving at road intersections: Safeguarding against collision, overly conservative behavior, and violation vehicles,” IEEE Trans. Ind. Electron., vol. 66, no. 4, pp. 3275–3286, 2019

    work page 2019

  4. [4]

    Extracting rules from autonomousvehicle-involved crashes by applying decision tree and association rule methods,

    M. T. Ashraf, K. Dey, S. Mishra, and M. T. Rahman, “Extracting rules from autonomousvehicle-involved crashes by applying decision tree and association rule methods,” Transp. Res. Rec. , vol. 2675, no. 11, pp. 522 – 533, 2021

    work page 2021

  5. [5]

    Safe reinforcement learning of lane change decision making with risk-fused constraint,

    Z. Li, L. Xiong, B. Leng, P. Xu, and Z. Fu, “Safe reinforcement learning of lane change decision making with risk-fused constraint,” in Proc. IEEE Intell. Transp. Syst. Conf. , 2023, pp. 1313–1319

    work page 2023

  6. [6]

    On-ramp merging for highway autonomous driving: An application of a new safety indicator in deep reinforcement learning,

    G. Li, W. Zhou, S. Lin, S. Li, and X. Qu, “On-ramp merging for highway autonomous driving: An application of a new safety indicator in deep reinforcement learning,” Automot. Innov., vol. 6, no. 3, pp. 453 – 465, 2023

    work page 2023

  7. [7]

    Double deep q-networks based game-theoretic equilibrium control of automated vehicles at autonomous intersection,

    H. Hu, D. Chu, J. Yin, and L. Lu, “Double deep q-networks based game-theoretic equilibrium control of automated vehicles at autonomous intersection,” Automot. Innov., vol. 7, no. 4, pp. 571 – 587, 2024

    work page 2024

  8. [8]

    Towards safe decision- making for autonomous vehicles at unsignalized intersections,

    K. Yang, S. Li, Y . Chen, D. Cao, and X. Tang, “Towards safe decision- making for autonomous vehicles at unsignalized intersections,” IEEE Trans. Veh. Technol., vol. 74, no. 3, pp. 3830–3842, 2025

    work page 2025

Show all 40 references
  1. [9]

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

    S. Gu, L. Yang, Y . Du, G. Chen, F. Walter, J. Wang, and A. Knoll, “A review of safe reinforcement learning: Methods, theories, and applications,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 12, pp. 11 216–11 235, 2024

    2024

  2. [10]

    Responsive safety in reinforce- ment learning by pid lagrangian methods,

    A. Stooke, J. Achiam, and P. Abbeel, “Responsive safety in reinforce- ment learning by pid lagrangian methods,” in Int. Conf. Mach. Learn. PMLR, 2020, pp. 9133–9143

    2020

  3. [11]

    Meta sac-lag: Towards deployable safe reinforcement learning via metagradient-based hyperparameter tuning,

    H. Honari, A. M. S. Enayati, M. G. Tamizi, and H. Najjaran, “Meta sac-lag: Towards deployable safe reinforcement learning via metagradient-based hyperparameter tuning,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. IEEE, 2024, pp. 619–626

    2024

  4. [12]

    Constrained policy optimization,

    J. Achiam, D. Held, A. Tamar, and P. Abbeel, “Constrained policy optimization,” in Int. Conf. Mach. Learn. PMLR, 2017, pp. 22–31

    2017

  5. [13]

    First order constrained optimiza- tion in policy space,

    Y . Zhang, Q. Vuong, and K. Ross, “First order constrained optimiza- tion in policy space,” Adv. Neural Inf. Process. Syst. , vol. 33, pp. 15 338–15 349, 2020

    2020

  6. [14]

    The unmanned vehicle on-ramp merging model based on am-mappo algorithm,

    Z. Shixin, P. Feng, J. Anni, Z. Hao, and G. Qiuqi, “The unmanned vehicle on-ramp merging model based on am-mappo algorithm,” Sci. Rep., vol. 14, no. 1, p. 19416, 2024

    2024

  7. [15]

    Learn zero-constraint-violation safe policy in model-free constrained rein- forcement learning,

    H. Ma, C. Liu, S. E. Li, S. Zheng, W. Sun, and J. Chen, “Learn zero-constraint-violation safe policy in model-free constrained rein- forcement learning,” IEEE Trans. Neural Netw. Learn, Syst. , vol. 36, no. 2, pp. 2327–2341, 2025

    2025

  8. [16]

    Synthesize efficient safety certificates for learning-based safe control using magnitude regularization,

    H. Zheng, H. Ma, S. Zheng, S. E. Li, and J. Wang, “Synthesize efficient safety certificates for learning-based safe control using magnitude regularization,” in IEEE Int. Conf. Robot. Autom. , 2024, pp. 545–551

    2024

  9. [17]

    Ensuring safety of learning-based motion planners using control barrier functions,

    X. Wang, “Ensuring safety of learning-based motion planners using control barrier functions,” IEEE Robot. Autom. Lett. , vol. 7, no. 2, pp. 4773–4780, 2022

    2022

  10. [18]

    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,” in Eur. Control Conf. (ECC) , 2019, pp. 3420–3431

    2019

  11. [19]

    Identify, estimate and bound the uncertainty of reinforcement learning for autonomous driving,

    W. Zhou, Z. Cao, N. Deng, K. Jiang, and D. Yang, “Identify, estimate and bound the uncertainty of reinforcement learning for autonomous driving,” IEEE Trans. Intell. Transport. Syst., vol. 24, no. 8, pp. 7932– 7942, 2023

    2023

  12. [20]

    Aleatoric and epistemic uncer- tainty in machine learning: An introduction to concepts and methods,

    E. H ¨ullermeier and W. Waegeman, “Aleatoric and epistemic uncer- tainty in machine learning: An introduction to concepts and methods,” Mach. Learn., vol. 110, no. 3, pp. 457–506, 2021

    2021

  13. [21]

    Safe control against uncertainty: A compre- hensive review of control barrier function strategies,

    S. Wang and S. Wen, “Safe control against uncertainty: A compre- hensive review of control barrier function strategies,” IEEE Syst. Man Cybern. Mag., vol. 11, no. 1, pp. 34–47, 2025

    2025

  14. [22]

    Ensemble quantile networks: Uncertainty-aware reinforcement learning with applications in au- tonomous driving,

    C.-J. Hoel, K. Wolff, and L. Laine, “Ensemble quantile networks: Uncertainty-aware reinforcement learning with applications in au- tonomous driving,” IEEE Trans. Intell. Transport. Syst., vol. 24, no. 6, pp. 6030–6041, 2023

    2023

  15. [23]

    Safe reinforcement learning in autonomous driving with epistemic uncertainty estimation,

    Z. Zhang, Q. Liu, Y . Li, K. Lin, and L. Li, “Safe reinforcement learning in autonomous driving with epistemic uncertainty estimation,” IEEE Trans. Intell. Transport. Syst., vol. 25, no. 10, pp. 13 653–13 666, 2024

    2024

  16. [24]

    Towards robust decision-making for autonomous driving on highway,

    K. Yang, X. Tang, S. Qiu, S. Jin, Z. Wei, and H. Wang, “Towards robust decision-making for autonomous driving on highway,” IEEE Trans. Veh. Technol., vol. 72, no. 9, pp. 11 251–11 263, 2023

    2023

  17. [25]

    Uncertainty-aware decision-making for autonomous driving at un- controlled intersections,

    X. Tang, G. Zhong, S. Li, K. Yang, K. Shu, D. Cao, and X. Lin, “Uncertainty-aware decision-making for autonomous driving at un- controlled intersections,” IEEE Trans. Intell. Transport. Syst. , vol. 24, no. 9, pp. 9725–9735, 2023

    2023

  18. [26]

    Ensemble deep learning: A review,

    M. A. Ganaie, M. Hu, A. K. Malik, M. Tanveer, and P. N. Suganthan, “Ensemble deep learning: A review,” Eng. Appl. Artif. Intell., vol. 115, p. 105151, 2022

    2022

  19. [27]

    Dsac: Distribu- tional soft actor critic for risk-sensitive reinforcement learning,

    X. Ma, L. Xia, Z. Zhou, J. Yang, and Q. Zhao, “Dsac: Distribu- tional soft actor critic for risk-sensitive reinforcement learning,” arXiv preprint arXiv:2004.14547, 2020

    2004 arXiv

  20. [28]

    Robust estimation of a location parameter,

    P. J. Huber, “Robust estimation of a location parameter,” in Break- throughs in statistics: Methodology and distribution . Springer, 1992, pp. 492–518

    1992

  21. [29]

    High-order control barrier functions,

    W. Xiao and C. Belta, “High-order control barrier functions,” IEEE Trans. Autom. Control., vol. 67, no. 7, pp. 3655–3662, 2022

    2022

  22. [30]

    Discrete-time control barrier function: High-order case and adaptive case,

    Y . Xiong, D.-H. Zhai, M. Tavakoli, and Y . Xia, “Discrete-time control barrier function: High-order case and adaptive case,” IEEE Trans. Cybern., vol. 53, no. 5, pp. 3231–3239, 2023

    2023

  23. [31]

    Conditional value-at-risk for general loss distributions,

    R. Rockafellar and S. Uryasev, “Conditional value-at-risk for general loss distributions,” J. Bank. Financ. , vol. 26, no. 7, pp. 1443–1471, 2002

    2002

  24. [32]

    Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation,

    A. Rame and M. Cord, “Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation,” in Proc. Int. Conf. Learn. Represent., 2021

    2021

  25. [33]

    Deep exploration via bootstrapped dqn,

    I. Osband, C. Blundell, A. Pritzel, and B. Van Roy, “Deep exploration via bootstrapped dqn,” Adv. Neural Inf. Process. Syst. , vol. 29, 2016

    2016

  26. [34]

    Randomized prior functions for deep reinforcement learning,

    I. Osband, J. Aslanides, and A. Cassirer, “Randomized prior functions for deep reinforcement learning,” Adv. Neural Inf. Process. Syst. , vol. 31, 2018

    2018

  27. [35]

    Rajamani, Vehicle dynamics and control

    R. Rajamani, Vehicle dynamics and control . Springer Science & Business Media, 2011

    2011

  28. [36]

    High-order control barrier functions: In- sights and a truncated taylor-based formulation,

    J. Xu and B. Alrifaee, “High-order control barrier functions: In- sights and a truncated taylor-based formulation,” arXiv preprint arXiv:2503.15014, 2025

    2025 arXiv

  29. [37]

    CasADi – A software framework for nonlinear optimization and optimal control,

    J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl, “CasADi – A software framework for nonlinear optimization and optimal control,” Mathematical Programming Computation , 2018

    2018

  30. [38]

    An environment for autonomous driving decision- making,

    E. Leurent, “An environment for autonomous driving decision- making,” https://github.com/eleurent/highway-env, 2018

    2018

  31. [39]

    Congested traffic states in empirical observations and microscopic simulations,

    M. Treiber, A. Hennecke, and D. Helbing, “Congested traffic states in empirical observations and microscopic simulations,” Phys. Rev. E, pp. 1805–1824, Jul 2002

    2002

  32. [40]

    Risk-aware reinforcement learning for autonomous driving: Improving safety when driving through intersection,

    B. Leng, R. Yu, W. Han, L. Xiong, Z. Li, and H. Huang, “Risk-aware reinforcement learning for autonomous driving: Improving safety when driving through intersection,” arXiv preprint arXiv:2503.19690, 2025

    2025 arXiv

Pith tools

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