The reviewed record of science sign in
Pith

arxiv: 2607.05180 · v1 · pith:EPRSMREB · submitted 2026-07-06 · cs.RO · cs.CV· cs.SY· eess.SY

VLM-CASE: Vision-Language Model Enabled Context-Adaptive Safety Envelopes for Anticipatory Safe Autonomous Driving

Tianjia Yang , Ke Li , Ruwen Qin , Xianbiao Hu This is my paper

Reviewed by Pith2026-07-08 00:48 UTCglm-5.2pith:EPRSMREBopen to challenge →

classification cs.RO cs.CVcs.SYeess.SY
keywords vision-language modelsautonomous drivingresponsibility-sensitive safetysafety envelopemodel predictive controladverse driving conditionsfriction estimationcontext-adaptive control
0
0 comments X

The pith

VLM sets the safety envelope, not just the driving plan

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

This paper proposes VLM-CASE, a framework where a vision-language model's scene understanding sets the vehicle's formal safety limits — not just its driving decisions. A LoRA-fine-tuned VLM classifies road surface and visibility from camera images every 2 seconds, producing two parameters: tire-road friction (μ) and forward observability (o_f). These parameters feed a context-adaptive safety envelope (CASE) derived from Responsibility-Sensitive Safety (RSS) principles, which couples braking and steering through a shared friction budget and widens following distance as perception degrades. An MPC controller then optimizes freely within this envelope at 20 Hz, never blocked by the slower VLM. The key architectural insight is the routing: scene understanding flows into the safety boundary (θ → U_safe), not into the cost function or action directly, so the formal safety guarantee is preserved while the vehicle gains anticipatory, human-like caution. In 198 closed-loop CARLA simulation runs, VLM-CASE-MPC completed 100% of trials versus 52% for Base MPC, 76% for VLM-MPC, and 82% for Fixed-Envelope MPC, with ablations showing friction and visibility adaptations address complementary failure modes.

Core claim

The central discovery is that routing VLM-derived scene understanding into the vehicle's formal safety envelope parameters — rather than into driving decisions alone — preserves the formal safety guarantee while enabling anticipatory adaptation. The context-adaptive safety envelope (CASE) is the object that carries this: it takes friction μ and observability o_f from the VLM's scene classification and uses them to set three coupled constraint sets (friction circle, visibility-aware following distance, lane-keeping drift bound) that bound all admissible actions. The shared friction budget is the mechanism that captures how adverse conditions couple braking and steering — lateral demand in a曲线

What carries the argument

Context-Adaptive Safety Envelope (CASE): an intersection of three constraint sets — U_fric (friction circle coupling braking and steering), U_lon (RSS-derived following distance with visibility margin), and U_lat (lane-keeping drift bound) — parametrized by θ = [μ, o_f] from the VLM. The VLM runs asynchronously (2 s update) while MPC runs at 20 Hz, so semantic inference never blocks real-time control.

Load-bearing premise

The formal safety guarantee depends on the VLM correctly classifying road surface and visibility conditions, because the envelope's friction coefficient and following-distance margins are set from a discrete lookup table keyed on the VLM's output. If the VLM reports 'dry' on a snow-covered road, the envelope assumes more grip than exists and the guarantee silently fails. The 98.3% accuracy was measured on CARLA test images; real-world distribution shift would likely lower it,

What would settle it

A single VLM misclassification at the wrong moment — e.g., reporting dry (μ=0.8) on snow (actual μ=0.2) — would set the RSS stopping distance too short and void the formal guarantee without any detectable alarm.

Figures

Figures reproduced from arXiv: 2607.05180 by Ke Li, Ruwen Qin, Tianjia Yang, Xianbiao Hu.

Figure 1
Figure 1. Figure 1: Overview of VLM-CASE, an anticipatory, safety-guaranteed autonomous driving framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the VLM context inference pipeline. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Structure of the context-classification prompt [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Front-camera frames for the evaluated conditions (road surface, weather, time of day, illumination). [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Run outcomes in lead-braking experiments by condition and controller. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trajectories under the snow condition (C3) in lead-braking experiments. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ego speed profiles in no-lead experiments on the snow surface. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Following gap versus route completion under varying visibility. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
read the original abstract

Adverse driving conditions, such as bad weather, remain a principal barrier to autonomous driving because they degrade two things at once: what the vehicle can perceive and what it can physically do. Human drivers cope by anticipation, reasoning about the scene and re-budgeting speed, following distance, and steering before grip or sight is lost, whereas current autonomous driving systems at best react after the fact. This paper proposes VLM-CASE, a framework that gives an autonomous vehicle this anticipatory capacity while keeping its motion bounded by a formal safety model at all times. A vision-language model (VLM), fine-tuned with low-rank adaptation (LoRA), reasons about the scene from the front-camera image and reports the road surface and visibility conditions. This output parametrizes a context-adaptive safety envelope (CASE), derived from physical limits and the guarantees of responsibility-sensitive safety, that couples braking and steering through a shared friction budget. A model predictive controller then drives freely within the envelope, while the VLM runs asynchronously so it never blocks the real-time control loop. We validate the framework in closed-loop CARLA simulation on tasks that demand both lateral and longitudinal control, across a range of weather, road-surface, and lighting conditions. The resulting controller, VLM-CASE-MPC, completes all trials, outperforming a conventional MPC baseline and a state-of-the-art VLM-integrated controller. Ablations confirm that the gains come from context adaptation, with the friction and visibility adaptations proving complementary. Furthermore, the framework is controller-agnostic and pairs with almost any low-level controller, offering a promising direction for safe autonomous driving. The dataset and supplementary materials for VLM-CASE are available at https://github.com/ytj254/VLM-CASE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 8 minor

Summary. The paper proposes VLM-CASE, a framework in which a vision-language model (VLM) fine-tuned with LoRA classifies road surface and visibility conditions from a front-camera image, and these classifications are mapped to physical parameters (friction coefficient μ, forward observability o_f) that parametrize a context-adaptive safety envelope derived from RSS principles. The envelope couples longitudinal and lateral safety through a shared friction budget (Eq. 10) and widens following distance under degraded visibility (Eq. 13). An MPC controller then operates within this envelope at 20 Hz while the VLM runs asynchronously every 2 seconds. The system is evaluated in CARLA simulation across 198 runs spanning three experiment groups (no-lead, constant-lead, lead-braking) with ablations isolating friction and visibility adaptations. VLM-CASE-MPC achieves 100% success across all runs, compared to 52% for Base MPC, 76% for VLM-MPC, and 82% for Fixed-Envelope MPC.

Significance. The paper's core architectural idea — routing VLM-derived scene understanding into the formal safety envelope rather than only into the planning objective — is a genuine and well-motivated contribution. The RSS-derived envelope with friction-circle coupling (Eq. 10–11) is physically grounded, and the re-derivation of lateral RSS against lane boundaries (Eq. 15–17) with braking-dependent lateral drift is a non-trivial extension. The asynchronous architecture cleanly separates second-scale semantic inference from millisecond-scale control. The experimental design is thorough: 198 runs, three experiment groups isolating friction and visibility, ablations (Table 5), and multiple baselines including a reproduction of VLM-MPC (Long et al., 2026). The 100% success rate against 50–82% baselines on adverse conditions, with the ablation confirming complementarity of the two adaptations, is a compelling empirical result. The code and dataset are publicly available, which supports reproducibility.

major comments (1)
  1. The central claim is that the framework 'preserves a formal safety guarantee' (§1, §3.1, §6). However, the guarantee chain is conditional on the VLM correctly classifying the road surface: the RSS-derived envelope uses b(μ) = min{μg, b_max} (Eq. 12) and the friction circle (Eq. 10), both of which assume the VLM-reported μ upper-bounds the actual friction. If the VLM misclassifies snow as dry (μ=0.8 instead of 0.2), the envelope assumes 4× the available grip, the RSS stopping distance is too short, and the friction circle permits accelerations the surface cannot support. The paper's own Fixed-Envelope MPC baseline demonstrates this: on snow (C3, Table 2), it achieves only 50% success with lane departures and minimum boundary clearance of 0.08 m — exactly the failure mode that would result from a surface misclassification. The paper does not analyze the safety implications of VLM misclassi
minor comments (8)
  1. §3.3.2, Eq. (13): The observability margin Δ_lon uses parameters d_0, τ, and α, but their values are not reported in the main text or tables. Please include the calibrated values.
  2. §3.3.3, Eq. (15): The lateral drift formula uses a_resp_y = sqrt((μg)² - a_x²), which couples lateral and longitudinal through the friction budget. This is a key design choice. Please clarify whether a_x is the commanded or achieved acceleration, and how feasibility is ensured when braking demand leaves no residual lateral authority.
  3. Table 2: The note states lateral metrics are omitted for VLM-MPC because it has no lateral control. However, VLM-MPC is supplemented with a basic waypoint-following lateral controller (§4.2). Please report lateral metrics for VLM-MPC or clarify why the supplemented controller does not perform lane keeping.
  4. Table 3, N3 snow row: The mean lateral error for VLM-CASE-MPC is 0.20 m, substantially higher than the 0.02–0.03 m on dry/wet. The paper attributes this to the full-route scoring effect, but the boundary clearance (0.18 m) is also much lower than on dry/wet (0.65–0.72 m). Please discuss whether this represents a genuine degradation in tracking quality on snow or a deliberate speed-position tradeoff.
  5. §4.3: The fine-tuning dataset (10,560 frames) is collected in CARLA and evaluated on CARLA test images. The 98.3% accuracy (Table 1) thus reflects in-distribution performance. Please note this limitation explicitly near the accuracy claim.
  6. §3.4.2, Eq. (20): The cornering penalty uses parameters w_c and η, but their values are not reported. Please include them.
  7. Figure 4: The condition labels are small and difficult to read. Consider enlarging or providing a higher-resolution version.
  8. The paper uses 'Qwen3-VL' as the backbone model (§4.3). At the time of this manuscript, the publicly known model is Qwen2-VL. Please verify the model name and version.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying the most important conceptual gap in the paper. The referee is correct that the safety guarantee is conditional on the VLM's classification being correct, and that the paper does not adequately analyze the consequences of misclassification. We agree this must be addressed.

read point-by-point responses

  1. Referee: The central claim is that the framework 'preserves a formal safety guarantee' (§1, §3.1, §6). However, the guarantee chain is conditional on the VLM correctly classifying the road surface: the RSS-derived envelope uses b(μ) = min{μg, b_max} (Eq. 12) and the friction circle (Eq. 10), both of which assume the VLM-reported μ upper-bounds the actual friction. If the VLM misclassifies snow as dry (μ=0.8 instead of 0.2), the envelope assumes 4× the available grip, the RSS stopping distance is too short, and the friction circle permits accelerations the surface cannot support. The paper's own Fixed-Envelope MPC baseline demonstrates this: on snow (C3, Table 2), it achieves only 50% success with lane departures and minimum boundary clearance of 0.08 m — exactly the failure mode that would result from a surface misclassification. The paper does not analyze the safety implications of VLM misclassi

    Authors: The referee is correct on all counts. The safety guarantee in VLM-CASE is conditional: it holds given that the VLM-reported friction coefficient μ upper-bounds the true available friction. If the VLM misclassifies snow as dry, the envelope assumes μ=0.8 against a true μ=0.2, and the guarantee breaks — the RSS stopping distance is too short and the friction circle permits accelerations the surface cannot support. The Fixed-Envelope MPC baseline on snow (C3, Table 2) is precisely this failure mode: it assumes dry grip on a snow surface and achieves only 50% success with lane departures and 0.08 m minimum clearance. A VLM misclassification would produce the same result. We will make the following revisions: (1) Qualify the safety guarantee language throughout the paper (§1, §3.1, §6) to state explicitly that the guarantee is conditional on the VLM classification being a correct upper bound on the true friction, rather than claiming an unconditional guarantee. (2) Add a dedicated subsection in §6 analyzing the misclassification failure mode, including: the measured 98.3% all-fields accuracy (Table 1) implies that approximately 1.7% of frames carry an incorrect context; the asynchronous 2-second update means a misclassification persists for multiple control steps; and the specific failure mode for each misclassification type (e.g., snow→dry produces the Fixed-Envelope failure mode; dry→snow produces over-conservative but safe behavior). (3) Discuss mitigations that preserve a conditional guarantee under bounded classification error: conservative friction assignment (e.g., reporting the lower bound of a confidence interval rather than a point estimate), fallback to a worst-case μ when VLM confidence is low, and cross-checking VLM-derived μ against runtime friction estimators. revision: yes

Circularity Check

0 steps flagged

No significant circularity; the derivation is self-contained and the one self-citation is not load-bearing.

full rationale

The paper's central claim—that routing VLM-derived scene understanding into the safety envelope preserves a formal safety guarantee—is derived from RSS (Shalev-Shwartz et al., 2017, an external citation) and standard vehicle dynamics (the friction circle, Eq. 10). The envelope equations (Eqs. 9-17) follow from RSS worst-case braking arguments with context-dependent parameters substituted in. The friction values (µ_dry=0.8, µ_wet=0.4, µ_snow=0.2) are calibrated to CARLA surfaces and then the system is validated in CARLA on those same surfaces; this is a fitted-input concern (the specific performance numbers depend on the calibration matching the simulator), but the paper transparently states 'The three values are calibrated in CARLA' (Section 3.2.3) and acknowledges 'transferring the framework to real vehicles will require...validating the friction calibration on real surfaces' (Section 6). This is a generalization limitation, not circularity—the derivation chain itself (RSS → context-dependent envelope → MPC) is not tautological. The one self-citation (Yang et al., 2025, SimLKAS) appears in the literature review for lane-keeping verification and is not load-bearing for any derivation step. No equation reduces to its own inputs by construction, no prediction is a renamed fit, and no uniqueness theorem is invoked from the authors' own prior work. The 100% success rate is an empirical result, not a 'prediction' derived from a fitted parameter. The conditional nature of the safety guarantee (it holds only if the VLM classifies correctly) is a correctness/robustness concern, not a circularity concern—the guarantee is not defined in terms of its own conclusion. Score 1 reflects the minor fitted-input issue that is transparently disclosed but does not undermine the derivation's logical structure.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The axiom ledger captures the load-bearing assumptions and free parameters of the VLM-CASE framework. The most significant concern is that several parameters (visibility margin d0/τ/α, MPC weights, visibility scores) are not explicitly stated in the paper, and the friction values are calibrated to the simulation environment in which they are validated.

free parameters (7)
  • µ_dry, µ_wet, µ_snow = 0.8, 0.4, 0.2
    Friction coefficients calibrated to CARLA surfaces (Section 3.2.3, Section 4.1). These are lookup-table values, not independently measured.
  • d0, τ, α (visibility margin) = Not stated in paper
    Parameters of the observability margin ∆_lon (Eq. 13). d0 and τ set base and speed-proportional buffer; α controls nonlinearity. Values and calibration method not provided.
  • ρ_lon, ρ_lat (response times) = Not explicitly stated
    Longitudinal and lateral response times in RSS-derived constraints (Eqs. 12, 15). Standard RSS parameters but specific values not tabulated.
  • b_max, b_L (braking limits) = Not explicitly stated
    Mechanical braking limit and lead vehicle braking assumption in Eq. 12. Values not tabulated.
  • MPC weights (w_y, w_ψ, w_v, w_s, w_a, w_δ, w_Δa, w_Δδ, w_c, β, η) = Not explicitly stated
    Stage cost weights and cornering penalty parameters (Eqs. 19-20). Values not tabulated in the paper.
  • Visibility scores s_w, s_d, s_l = Not explicitly stated
    Per-field visibility scores in [0,1] for weather, time of day, illumination (Eq. 8). Specific values for each category not provided.
  • LoRA rank r, scaling α = r=16, α=32
    LoRA adaptation parameters (Section 4.3). Applied to query and value projections of last 12 transformer layers.
axioms (5)
  • domain assumption RSS proper-response framework: a vehicle executing the proper response never causes a collision, given worst-case assumptions about other vehicles' accelerations
    Invoked in Section 3.3.2 to derive the longitudinal safety constraint. The guarantee holds only if the worst-case acceleration bounds (including µ) are correct.
  • standard math Friction circle: total tire force is bounded by √(a_x² + a_y²) ≤ µg
    Invoked in Eq. 10-11. Standard vehicle dynamics model.
  • domain assumption Kinematic bicycle model adequately represents vehicle dynamics for control purposes
    Used as the MPC prediction model (Eq. 18). Valid at moderate speeds but degrades at high speeds or near handling limits.
  • domain assumption Lead vehicle maintains constant speed over the MPC prediction horizon
    Stated in Section 3.4.1. Simplifies the prediction but is violated during lead-braking scenarios, though the safety envelope compensates.
  • ad hoc to paper VLM classification of scene conditions is sufficiently accurate to set safety parameters
    The entire framework depends on the VLM correctly classifying road surface and visibility. The 98.3% accuracy is on CARLA test images only (Table 1). No analysis of safety consequences of misclassification is provided.
invented entities (1)
  • Forward observability o_f no independent evidence
    purpose: Composite visibility index parametrizing the longitudinal safety margin
    Defined in Eq. 8 as an equal-weight mean of per-field visibility scores. The scores themselves are not independently calibrated against any external visibility metric. The margin formula (Eq. 13) using o_f has free parameters whose values are not stated. No falsifiable prediction outside the CARLA environment is provided.

pith-pipeline@v1.1.0-glm · 23253 in / 4725 out tokens · 167459 ms · 2026-07-08T00:48:38.517231+00:00 · methodology

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

34 extracted references · 34 canonical work pages · 4 internal anchors

  1. [1]

    IEEE Transactions on Control Systems Technology 31, 2744–2759

    Control Barrier Functions and Input-to-State Safety With Application to Automated Vehicles. IEEE Transactions on Control Systems Technology 31, 2744–2759. doi:10.1109/TCST.2023.3286090. Ames, A.D., Grizzle, J.W., Tabuada, P.,

    work page doi:10.1109/tcst.2023.3286090 2023

  2. [2]

    6271–6278

    Control barrier function based quadratic pro- grams with application to adaptive cruise control, in: 53rd IEEE Conference on Decision and Control, pp. 6271–6278. doi:10.1109/CDC.2014.7040372. 23 Ames, A.D., Xu, X., Grizzle, J.W., Tabuada, P.,

    work page doi:10.1109/cdc.2014.7040372 2014

  3. [3]

    IEEE Transactions on Automatic Control 62, 3861–3876

    Control Barrier Function Based Quadratic Programs for Safety Critical Systems. IEEE Transactions on Automatic Control 62, 3861–3876. doi:10.1109/TAC.2016.2638961. Bijelic, M., Gruber, T., Mannan, F., Kraus, F., Ritter, W., Dietmayer, K., Heide, F.,

    work page doi:10.1109/tac.2016.2638961 2016

  4. [4]

    2082–2087

    Perception-Aware Chance-Constrained Model Predictive Control for Uncertain Environments, in: 2021 American Control Con- ference (ACC), pp. 2082–2087. doi:10.23919/ACC50511.2021.9483203. Boyapati, R.K., Eddy, M.M., Seitz, T.,

    work page doi:10.23919/acc50511.2021.9483203 2021

  5. [5]

    Transportation Re- search Record 2674, 662–673

    Evaluation and Optimization of Responsibility- Sensitive Safety Models on Autonomous Car-Following Maneuvers. Transportation Re- search Record 2674, 662–673. doi:10.1177/0361198120948507. Dean, S., Matni, N., Recht, B., Ye, V.,

    work page doi:10.1177/0361198120948507

  6. [6]

    Artificial Intelligence for Transportation 6, 100061

    From clear to adverse: A survey of all-weather LiDAR semantic segmentation for autonomous driving. Artificial Intelligence for Transportation 6, 100061. doi:10.1016/j.ait.2026.100061. He, S., Zeng, J., Zhang, B., Sreenath, K.,

    work page doi:10.1016/j.ait.2026.100061 2026

  7. [7]

    Rule-Based Safety-Critical Control De- sign using Control Barrier Functions with Application to Autonomous Lane Change, in: 2021 American Control Conference (ACC), pp. 178–185. doi:10.23919/ACC50511.2021. 9482848. Heinzler, R., Piewak, F., Schindler, P., Stork, W.,

    work page doi:10.23919/acc50511.2021 2021

  8. [8]

    IEEE Robotics and Automation Letters 5, 2514–2521

    CNN-Based Lidar Point Cloud De-Noising in Adverse Weather. IEEE Robotics and Automation Letters 5, 2514–2521. doi:10.1109/LRA.2020.2972865. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.,

    work page doi:10.1109/lra.2020.2972865 2020

  9. [9]

    EMMA: End-to-End Multimodal Model for Autonomous Driving

    EMMA: End-to-End Multimodal Model for Autonomous Driving.arXiv:2410.23262. 24 Jiang, S., Huang, Z., Qian, K., Luo, Z., Zhu, T., Zhong, Y., Tang, Y., Kong, M., Wang, Y., Jiao, S.,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Friction 5, 123–146

    A technical survey on tire-road friction esti- mation. Friction 5, 123–146. doi:10.1007/s40544-017-0151-0. Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E., Lam, G., Sanketi, P.,

    work page doi:10.1007/s40544-017-0151-0

  11. [11]

    OpenVLA: An Open-Source Vision-Language-Action Model

    Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246arXiv:2406.09246. Koopman, P., Osyk, B., Weast, J.,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Robotics and Autonomous Systems 173, 104621

    Path tracking of varying-velocity 4WS autonomous vehicles un- der tire force friction ellipse constraints. Robotics and Autonomous Systems 173, 104621. doi:10.1016/j.robot.2024.104621. Li, X., Liu, C., Chen, B., Jiang, J.,

    work page doi:10.1016/j.robot.2024.104621 2024

  13. [13]

    IEEE Transac- tions on Intelligent Transportation Systems 23, 20798–20809

    Robust Adaptive Learning-Based Path Tracking Control of Autonomous Vehicles Under Uncertain Driving Environments. IEEE Transac- tions on Intelligent Transportation Systems 23, 20798–20809. doi:10.1109/TITS.2022. 3176970. Liu, J., Liu, Y., Wang, L., Qu, X.,

    work page doi:10.1109/tits.2022 2022

  14. [14]

    IEEE Transactions on Intelligent Transportation Systems 26, 12819–12833

    A Survey and Comprehensive Taxonomy of Tire-Road Adhesion Coefficient Estimation for Intelligent Vehicles. IEEE Transactions on Intelligent Transportation Systems 26, 12819–12833. doi:10.1109/TITS.2025.3565542. Liu, S., Wang, X., Hassanin, O., Xu, X., Yang, M., Hurwitz, D., Wu, X.,

    work page doi:10.1109/tits.2025.3565542 2025

  15. [15]

    Transportation Research Part C: Emerging Technologies 125, 103037

    Calibration and evaluation of responsibility-sensitive safety (RSS) in automated vehicle performance during cut-in scenarios. Transportation Research Part C: Emerging Technologies 125, 103037. doi:10.1016/j.trc.2021.103037. Long, K., Shi, H., Liu, J., Xiao, C., Li, X.,

    work page doi:10.1016/j.trc.2021.103037 2021

  16. [16]

    Lyu, Y., Luo, W., Dolan, J.M.,

    doi:10.1016/j.trc.2025.105487. Lyu, Y., Luo, W., Dolan, J.M.,

    work page doi:10.1016/j.trc.2025.105487 2025

  17. [17]

    Adaptive Safe Merging Control for Heterogeneous Au- tonomous Vehicles using Parametric Control Barrier Functions, in: 2022 IEEE Intelligent Vehicles Symposium (IV), pp. 542–547. doi:10.1109/IV51971.2022.9827329. Neumeister, D.M., Pape, D.B.,

    work page doi:10.1109/iv51971.2022.9827329 2022

  18. [18]

    Systems & Control Letters 196, 106021

    Robust responsibility-sensitive safety: Noise disturbed adaptive cruise control. Systems & Control Letters 196, 106021. doi:10.1016/j.sysconle.2025.106021. 25 Salvi, A., Weiss, G., Trapp, M., Oboril, F., Buerkle, C.,

    work page doi:10.1016/j.sysconle.2025.106021 2025

  19. [19]

    Safety implications of runtime adaptation to changing operating conditions, in: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), IEEE. pp. 2444–2449. Sapkota, R., Cao, Y., Roumeliotis, K.I., Karkee, M.,

    work page 2022

  20. [20]

    arXiv preprint arXiv:2505.04769 arXiv:2505.04769

    Vision-language-action mod- els: Concepts, progress, applications and challenges. arXiv preprint arXiv:2505.04769 arXiv:2505.04769. Sezgin, F., Vriesman, D., Steinhauser, D., Lugner, R., Brandmeier, T.,

    work page arXiv

  21. [21]

    Safe Au- tonomous Driving in Adverse Weather: Sensor Evaluation and Performance Monitoring, in: 2023 IEEE Intelligent Vehicles Symposium (IV), pp. 1–6. doi:10.1109/IV55152.2023. 10186596. Shalev-Shwartz, S., Shammah, S., Shashua, A.,

    work page doi:10.1109/iv55152.2023 2023

  22. [22]

    On a Formal Model of Safe and Scalable Self-driving Cars

    On a formal model of safe and scalable self-driving cars. doi:10.48550/arXiv.1708.06374. Song, Z., Liu, L., Jia, F., Luo, Y., Jia, C., Zhang, G., Yang, L., Wang, L.,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1708.06374

  23. [23]

    IEEE Transactions on Intelligent Transportation Systems 25, 15407–15436

    Robustness- Aware 3D Object Detection in Autonomous Driving: A Review and Outlook. IEEE Transactions on Intelligent Transportation Systems 25, 15407–15436. doi:10.1109/TITS. 2024.3439557. Tian, X., Gu, J., Li, B., Liu, Y., Wang, Y., Zhao, Z., Zhan, K., Jia, P., Lang, X., Zhao, H.,

    work page doi:10.1109/tits 2024

  24. [24]

    Vehicle System Dynamics 62, 347–371

    Friction-Adaptive Stochastic Nonlinear Model Predictive Control for Autonomous Vehicles. Vehicle System Dynamics 62, 347–371. doi:10.1080/00423114.2023.2219791,arXiv:2305.03798. Wang, Y., Hu, J., Wang, F., Dong, H., Yan, Y., Ren, Y., Zhou, C., Yin, G.,

    work page doi:10.1080/00423114.2023.2219791 2023

  25. [25]

    Xu, X., Wang, X., Wu, X., Hassanin, O., Chai, C.,

    doi:10.1186/s10033-021-00675-z. Xu, X., Wang, X., Wu, X., Hassanin, O., Chai, C.,

    work page doi:10.1186/s10033-021-00675-z

  26. [26]

    Transportation Research Part C: Emerging Technologies 123, 102988

    Calibration and evaluation of the Responsibility-Sensitive Safety model of autonomous car-following maneuvers using naturalistic driving study data. Transportation Research Part C: Emerging Technologies 123, 102988. doi:10.1016/j.trc.2021.102988. Xu, Z., Zhang, Y., Xie, E., Zhao, Z., Guo, Y., Wong, K.Y.K., Li, Z., Zhao, H.,

    work page doi:10.1016/j.trc.2021.102988 2021

  27. [27]

    IEEE Robotics and Automation Letters 9, 8186–8193

    DriveGPT4: Interpretable End-to-End Autonomous Driving Via Large Language Model. IEEE Robotics and Automation Letters 9, 8186–8193. doi:10.1109/LRA.2024.3440097. Yang, T., Ding, Y., Li, K., Pan, Y., Qin, R., Yin, Z., Hu, X.,

    work page doi:10.1109/lra.2024.3440097 2024

  28. [28]

    Journal of Intelligent Transportation Systems , 1–25doi:10.1080/15472450

    SimLKAS: A simulation-based framework for the verification and validation of lane keeping assistance systems. Journal of Intelligent Transportation Systems , 1–25doi:10.1080/15472450. 2025.2559224. 26 Zang, S., Ding, M., Smith, D., Tyler, P., Rakotoarivelo, T., Kaafar, M.A.,

    work page doi:10.1080/15472450 2025

  29. [29]

    IEEE Vehicular Technology Magazine 14, 103–111

    The Impact of Adverse Weather Conditions on Autonomous Vehicles: How Rain, Snow, Fog, and Hail Affect the Performance of a Self-Driving Car. IEEE Vehicular Technology Magazine 14, 103–111. doi:10.1109/MVT.2019.2892497. Zhang, Y., Carballo, A., Yang, H., Takeda, K.,

    work page doi:10.1109/mvt.2019.2892497 2019

  30. [30]

    ISPRS Journal of Photogrammetry and Remote Sensing 196, 146–177

    Perception and sensing for autonomous vehicles under adverse weather conditions: A survey. ISPRS Journal of Photogrammetry and Remote Sensing 196, 146–177. doi:10.1016/j.isprsjprs.2022.12.021. Zhao, S., Zhang, J., Jiang, Y., He, C., Han, J.,

    work page doi:10.1016/j.isprsjprs.2022.12.021 2022

  31. [31]

    Mechanical Systems and Signal Processing 224, 112039

    Tire-Road friction coefficients adaptive estimation through image and vehicle dynamics integration. Mechanical Systems and Signal Processing 224, 112039. doi:10.1016/j.ymssp.2024.112039. Zhao, T., Guo, P., Wei, Y.,

    work page doi:10.1016/j.ymssp.2024.112039 2024

  32. [32]

    Mechanical Systems and Signal Processing 208, 111019

    Road friction estimation based on vision for safe au- tonomous driving. Mechanical Systems and Signal Processing 208, 111019. doi:10.1016/ j.ymssp.2023.111019. Zhou, X., Liu, M., Yurtsever, E., Zagar, B.L., Zimmer, W., Cao, H., Knoll, A.C.,

    work page arXiv 2023

  33. [33]

    IEEE Transactions on Intelligent Vehicles doi:10.1109/TIV.2024.3402136

    Vision Language Models in Autonomous Driving: A Survey and Outlook. IEEE Transactions on Intelligent Vehicles doi:10.1109/TIV.2024.3402136. Zhou, Z., Cai, T., Zhao, S.Z., Zhang, Y., Huang, Z., Zhou, B., Ma, J.,

    work page doi:10.1109/tiv.2024.3402136 2024

  34. [34]

    AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning

    Au- toVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning. arXiv preprint arXiv:2506.13757 arXiv:2506.13757. 27

    work page internal anchor Pith review Pith/arXiv arXiv