acceptodds
Under review as a conference paper at ICLR 2027

CoSA-Drive: Counterfactual Safety Authorization for Vision-Language-Action Models in Autonomous Driving

Abstract

Vision‑language‑action (VLA) models have shown strong capabilities in scene understanding, task reasoning, and trajectory planning for autonomous driving. However, driving VLAs rely on a decision chain that combines visual observations, temporal context, vehicle states, and task instructions. Corruption of any source can propagate through this chain and induce unsafe behavior. Existing defenses mainly detect abnormal inputs or restore corrupted observations. Detection only indicates that an input may be unreliable, while restoration does not establish that the resulting driving behavior is safe to execute. We therefore reformulate VLA defense as counterfactual behavior authorization and introduce CoSA‑Drive. The framework first audits multimodal evidence and constructs heterogeneous behavioral candidates using specialized defense experts. Its core authorizer, CERA, independently evaluates each candidate using runtime evidence and predicts its safety, intervention‑induced harm, and task utility. Calibrated confidence bounds authorize only candidates that satisfy the risk constraint. Defense experts can propose behaviors but cannot execute them directly. A deterministic DriveContract further checks traffic rules, task permissions, and operational constraints, and triggers replanning or a minimal‑risk maneuver when no candidate is admissible. Experiments across four driving‑VLA backbones and reactive CARLA environments show that CoSA‑Drive recovers 62.2–95.5% of attack‑induced degradation and improves the safety–utility trade‑off under multimodal attacks and distribution shifts.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
VLA Model, autonomous driving, multimodal attack, safety assurance

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dKPfOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoSA-Drive is a runtime defense framework for driving VLAs under visual, temporal, state, task, and cross-modal corruption. Different repair operators generate behavior candidates from a frozen driving model. CERA scores each candidate's safety, intervention-induced harm, and utility using runtime evidence and an ensemble uncertainty estimate. A deterministic DriveContract module then vetoes constraint violations, with replanning or minimum-risk fallback when necessary. The paper reports NAVSIM PDM increasing from 0.7503 under attack to 0.8399, with unsafe outcomes decreasing from 25.66% to 11.71%. Reactive CARLA experiments and tests on four VLA backends show additional recovery and safety benefits.

Critical strengths
  1. The behavior-level decision structure targets a real weakness in restoration-only defenses. Evaluating the consequences of multiple admissible interventions, while reserving final veto authority for a separate checker, is a sensible architecture (Sections 4.3–4.7). The NAVSIM comparison shows that visual restoration alone reduces PDM while candidate authorization improves both PDM and proxy safety (Table 2). Explicit utility and new-harm heads also address the risk that nominally conservative defenses merely stop the vehicle.

  2. The evaluation includes reactive outcomes and acknowledges selective-execution trade-offs. The CARLA studies compare restoration, shielding, component removal, and the full system; the appendix reports route-clustered uncertainty rather than treating correlated frames as independent observations (Tables 3, 5–6; Figure 11). Figure 11 supports improvements in red-light violations and any safety violation over Shield-only. Risk–coverage curves and target-backend calibration details further improve interpretability (Appendices D and G). These are substantial strengths relative to evaluating only image repair or trajectory similarity.

Critical weaknesses
  1. The operational problem and distinct contribution are insufficiently specified. Although the paper motivates evaluating behavior after input repair, the concrete authorization problem remains difficult to reconstruct: which runtime evidence is trustworthy, how candidate safety is labeled and predicted, and how learned authorization differs from the deterministic checker are not specified sufficiently. This makes it difficult to assess whether CERA introduces a substantive advance over conventional candidate scoring and safety filtering.

  2. Comparison with established defense methods is insufficient. The experiments include restoration-only, shielding-only, and internal ablation baselines, but do not clearly identify implementations of previously published defense or runtime-assurance methods. Consequently, the results establish benefits over the selected simplified configurations without demonstrating an advantage over existing approaches. Comparisons with matched access to trusted evidence, candidate behaviors, and fallback mechanisms are particularly important.

Questions for authors
  1. Can the authors provide a concrete end-to-end example specifying corrupted inputs, trusted runtime evidence, candidate trajectories, CERA scores, and the final DriveContract decision? What distinguishes the proposed authorization mechanism from standard learned candidate selection followed by safety filtering?

  2. Which experimental baselines implement previously published methods? How does CoSA-Drive compare with the closest applicable approach under matched information, candidate-generation, and fallback conditions?

Reviewer_eyuHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CoSA-Drive, a runtime assurance framework for driving vision-language-action models under multimodal corruption. It generates alternative behaviors by applying different interventions to the inputs of a frozen VLA policy. A learned module, CERA, assesses behavioral safety, intervention-induced harm, and task utility using runtime evidence and uncertainty estimates. DriveContract then applies deterministic checks before execution. The evaluation covers several driving VLA backends, NAVSIM, and reactive CARLA experiments.

Critical strengths

The paper presents a coherent framework for behavior-level authorization in driving VLA models under multimodal corruption. It explicitly considers intervention-induced harm alongside behavioral safety and task utility. This formulation offers a useful perspective on evaluating defenses through their effects on driving behavior.

Critical weaknesses
  1. The closed-loop results are difficult to reconcile with the accompanying statistical analysis. The effects in Figure 11 do not match the corresponding differences in Tables 3 and 16, and the paper does not explain whether this reflects different evaluation settings.
  2. The paper leaves important implementation details and baseline comparisons unclear. For example, it does not explain how the CARLA candidates in Figure 12 are generated or how they relate to Section 4.3. Similar gaps appear in Figures 8 and 10, where the selection rules for “Fixed expert” and “Risk-gated fixed” and their use of DriveContract and fallback are not fully specified.
Questions for authors
  1. Why do the improvements over Shield-only in Figure 11 differ from those implied by Tables 3 and 16?
  2. How are the CARLA candidates in Figure 12 generated, and how do they relate to the input interventions in Section 4.3?
  3. How are CERA’s safety labels assigned, and where does DriveContract obtain obstacle and traffic-signal information under corruption?
  4. How are “Fixed expert” and “Risk-gated fixed” in Figures 8 and 10 defined, and if other framework components are held fixed?
  5. What is CoSA-Drive’s end-to-end latency, and do the CARLA experiments account for this delay?
Reviewer_8Dv7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents CoSA-Drive, a runtime assurance framework that helps driving vision-language-action (VLA) models handle corrupted multimodal inputs safely. Rather than assuming repaired inputs are safe, it generates candidate behaviors through different interventions. It uses CERA, a learned, ensemble-based authorizer, to assess their safety, potential harm, and usefulness for the task. A deterministic DriveContract then checks each behavior against hard safety constraints before execution. Experiments with four VLA backends, NAVSIM, and reactive CARLA show promising results. On NAVSIM, CoSA-Drive improves PDM from 0.7503 to 0.8399 and cuts unsafe behaviors from 25.66% to 11.71%. In CARLA, it raises the Driving Score from 53.04 to 61.03 while reducing collisions from 25.00% to 11.67%. The paper also includes ablations, risk–coverage analysis, and statistical evaluations.

Critical strengths
  • The shift from input restoration to behavior-level authorization is well motivated. The examples show that even high-quality restoration may not improve downstream driving, highlighting a limitation of existing defenses. Separating candidate generation, learned authorization, and deterministic safety checks is a sensible approach for safety-critical systems.
  • The evaluation is thoughtfully designed, with separate training, calibration, and test splits, prespecified risk budgets, and risk–coverage curves. Route-clustered CARLA confidence intervals (Fig. 11), distinct ablation protocols, and a discussion of evaluation limits in Appendix G make the results easier to assess.
  • Experiments cover four different driving-VLA backends, open-loop planning, reactive closed-loop driving, and multimodal corruption. Ablations clarify how each component affects safety and task progress: removing DriveContract increases red-light violations from 11.67% to 38.33%.
Critical weaknesses
  • Closed-loop gains over the strongest baseline are less impressive than the headline results suggest. CoSA-Drive improves Driving Score over the attacked VLA, but its route-clustered gain over Shield-only (Fig. 11) is smaller and not statistically significant (+3.93, 95% CI [-0.34, 8.37], p = 0.148). The collision improvement is also insignificant (p = 0.076); only red-light and aggregate safety violations improve significantly. CoSA-Drive is not evaluated on clean inputs: under attack, it scores 61.03, above the undefended clean VLA’s 56.37. Because the clean policy already has substantial safety violations, it is unclear whether these gains come from recovering corrupted inputs or from general safety filtering. A clean-input CoSA-Drive baseline is needed to separate these effects. Table 2 also lacks a clean-VLA reference for NAVSIM.
  • Runtime feasibility and fallback costs are not reported. CoSA-Drive may require multiple VLA passes per decision, learned auditors, repair networks, and a five-model CERA ensemble, but the paper provides no latency, compute-overhead, or control-frequency data. Alpamayo 1.5 achieves only 46.5% normal-execution coverage (Table 4), meaning over half of attacked decisions are not approved for normal execution. Without information on delays, replanning, or minimum-risk maneuvers, the practical cost of this conservatism remains unclear.
  • Generalization and baseline comparisons need stronger support. Auditors, repair experts, and CERA are trained on evaluation-suite corruption data. Despite holding out test cases, the motion-integrity network's detection rate falls from 97.55% internally to 59.50% on held-out attacks (Table 9). Testing on a wholly unseen corruption family would better support generalization claims. Comparisons also focus more on the authors’ defenses than on established external baselines. Several DriveContract checks overlap with NAVSIM's PDM safety criteria, making it unclear whether NAVSIM’s improvement reflects broader safety gains or simply metric alignment.
  • Key evaluation details are unclear. H-ASR thresholds are underspecified, and Table 1 lacks confidence intervals for the threat model. Visual corruption yields 29.83% H-ASR despite slight improvements in mean ADE/FDE and PDM; the paper should clarify how harmful outcomes are counted. Task-channel results vary widely between ordinary malicious instructions and long-text attacks. Backend transfer still requires calibration on attacked data, a deployment requirement that warrants further discussion.
Questions for authors
  • Could the authors evaluate CoSA-Drive on clean CARLA inputs and report the clean-VLA baseline on NAVSIM? This would help distinguish attack recovery from general safety filtering.
  • Could the authors highlight the route-clustered confidence intervals in Fig. 11 and clarify which gains over Shield-only are statistically significant? Why does the aggregate Driving Score difference (5.58) differ from the paired route-clustered estimate (+3.93)?
  • What are the per-decision latency, number of VLA forward passes, and achieved control frequency? How often does CoSA-Drive use delayed execution, replanning, or minimum-risk maneuvers? Specifically, how does Alpamayo’s 46.5% authorization coverage affect task progress?
  • Could the authors test generalization by holding out an entire corruption family when training all components? This is especially relevant given the motion-integrity network's lower detection rate on held-out attacks.
  • How is H-ASR defined, including the metrics and thresholds for each information condition? Could the authors provide confidence intervals supporting the attack-validity claims in Section 3.1?
  • How much do DriveContract's checks overlap with NAVSIM's PDM submetrics? Could the authors report results using safety measures the contract does not directly enforce to better identify the source of the improvement?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.