Recoleta
GitHub
Source note

Lost in Fog: Sensor Perturbations Expose Reasoning Fragility in Driving VLAs

Authors
Abhinaw Priyadershi, Jelena Frtunikj
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionAutonomous DrivingSensor CorruptionChain Of CausationRuntime Monitoring
Open arXivMarkdown source

Summary

This paper tests Alpamayo R1, a 10B driving VLA, under camera noise, lighting shifts, and fog. It claims that changes in the model's chain-of-causation explanation are a strong warning signal for large trajectory shifts.

Problem

  • Driving VLAs can explain their plans, but those explanations may change when cameras degrade in fog, glare, darkness, or sensor noise.
  • This matters because a planner can output plausible text while its predicted path drifts, which is a safety issue for autonomous driving validation.

Approach

  • The authors evaluate Alpamayo R1 on 1,996 PhysicalAI autonomous-driving validation scenarios.
  • They apply 8 synchronized multi-camera perturbations: Gaussian noise at σ=10,30,50,70, brightness 0.4× and 1.6×, and fog α=0.3 and 0.7.
  • They run about 18,000 inference trials and compare clean versus perturbed outputs using ADE, ADE increase, L2 trajectory deviation, and exact-match CoC change rate.
  • They also run a CoC-suppression ablation with the same checkpoint and decoding settings except for token budget, so the causal claim is limited.

Results

  • On clean inputs, Alpamayo R1 reaches 2.00 m ADE versus 6.32 m for a constant-velocity baseline, a 68.3% reduction with p<10^-257 on 1,996 samples.
  • When the CoC explanation changes after perturbation, mean L2 trajectory deviation is 21.82 m versus 4.13 m when it stays unchanged, a 5.3× gap; median deviation is 15.39 m versus 2.16 m, a 7.1× gap.
  • Across all 15,968 attacked sample pairs, CoC change has point-biserial r=0.53 with trajectory deviation, Cohen's d=1.12, and p<10^-100; across 8 attack types the aggregate correlation is r=0.99.
  • Noise degradation rises nearly linearly over σ=10,30,50,70 with R²=0.957 and slope 0.0048 m ADE per σ unit; the >5 m trajectory-shift rate rises from 18.8% at σ=10 to 70.6% at σ=70.
  • Heavy noise σ=70 is the largest tested corruption: ADE rises to 2.30 m, ΔADE is +0.30 m, CoC changes in 52.7% of samples, and p=2×10^-17.
  • The CoC ablation reports 11.8% average ADE improvement when CoC generation is enabled across tested conditions, with p<0.0001; standard preprocessing defenses give only small, statistically weak relief.

Link