Inside the Reasoning Menu: Discrete Failure Bands in Driving VLAs Under Sensor Perturbations

Abhinaw Priyadershi, Jelena Frtunikj

Keywords: VLA robustness, autonomous driving, discrete reasoning, trajectory multimodality, sensor perturbation

TL;DR: VLA trajectory errors under sensor noise cluster into six discrete severity bands, not a smooth gradient; binary CoC-change detection outperforms continuous similarity metrics.

Abstract: Vision-Language-Action (VLA) planners produce natural-language explanations alongside driving trajectories. Prior work found that when sensor perturbations change the explanation, trajectory error spikes by $5.3\times$, but the magnitude of text change has near-zero predictive power. We investigate the source of this mismatch. Using Alpamayo R1, a 10B-parameter driving VLA, we analyze the L2 trajectory errors of 5,443 changed-explanation pairs drawn from 15,968 evaluation pairs spanning 8 perturbation types. A Gaussian Mixture Model identifies six discrete severity bands, with BIC selecting $k=6$ (a sharp improvement from $k=5$ to $k=6$) and the preference remaining stable across 20 restarts. An action-transition analysis reveals that 71% of explanation changes are surface rewrites within the same action category; action labels do not significantly predict which severity band a failure lands in ($p = 0.306$, $n=59$ genuine switches). The structure is consistent with sub-categorical organization: geometric context (distance, timing, object proximity) determines the band, not the action word. For VLA pipelines that use Chain-of-Causation (COC) text for downstream monitoring or fallback, binary COC-change detection is therefore a more useful trigger than Jaccard-style magnitude scoring.

Submission Number: 9