Title: Navigating the Fog: How Sensor Distortions Reveal the Weaknesses in Driving VLA Reasoning

Autonomous driving systems that offer interpretable plans must ensure that their explanations remain trustworthy even when real-world sensors degrade. This study introduces a controlled experiment to test the robustness of Vision-Language-Action (VLA) models in self-driving contexts. The research evaluated Alpamayo R1, a model with 10 billion parameters, across 1,996 distinct scenarios. These scenarios involved eight types of sensor perturbations, including Gaussian noise at four intensity levels, two extremes of lighting, and two degrees of fog density, resulting in approximately 18,000 inference trials.

The findings indicate that the stability of reasoning serves as a highly accurate marker for trajectory reliability. Specifically, when the Chain-of-Causation (CoC) explanations altered following a perturbation, the deviation in the vehicle’s path increased by a factor of 5.3, rising from an average of 4.1 meters to 21.8 meters. This correlation was strong, with a coefficient of $r!=!0.99$ across various attack types and a per-sample Pearson correlation of $r_{pb}!=!0.53$ (Cohen's $d!=!1.12$).

Furthermore, a controlled ablation study suggests that activating CoC generation leads to better trajectory accuracy, improving performance by an average of 11.8% across all conditions ($p < 0.0001$) under comparable inference settings. Within the tested noise spectrum ($\sigma \in {10, 30, 50, 70}$), performance degradation followed an approximately linear trend ($R^2!=!0.957$). Standard preprocessing techniques offered only minimal protection against these issues. Consequently, the study positions CoC consistency as a measurable proxy for planning safety and advocates for reasoning-based runtime monitoring to enhance the security of VLA deployments.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC