Does Visual Degradation Amplify Instruction Sensitivity in Vision-Language-Action Models? An Empirical Study with OpenVLA

Jihwan Woo

Keywords: Vision-Language-Action Models, Instruction Sensitivity, Visual Degradation Robustness, Harsh-Environment Deployment, Robot Manipulation

TL;DR: Visual degradation doesn't uniformly amplify VLA instruction sensitivity — occlusion causes model collapse, low lighting suppresses it, and when SNR drops below 1, the model can no longer meaningfully follow instructions.

Abstract: Vision-Language-Action (VLA) models deployed in harsh environments face compounded challenges: degraded visual inputs and natural variation in operator instructions. This paper investigates whether visual degradation amplifies instruction sensitivity, using OpenVLA-7B as a case study. Three manipulation tasks are evaluated under four degradation conditions (low lighting, Gaussian noise, fog, partial occlusion) at three intensity levels. The effect is degradation-type-dependent. Under occlusion, the model's action outputs become highly stochastic (noise floor $= 0.250$), and instruction content has diminishing influence — a pattern consistent with model collapse rather than amplified sensitivity. Low lighting suppresses sensitivity ($F = 15.11$, $p < 0.001$), while noise and fog show no significant effect. The signal-to-noise ratio decreases from $3.0\times$ (clean) to $0.7\times$ (severe), crossing the $\text{SNR} = 1$ threshold. We propose this as a deployment criterion: if the ratio of instruction-induced variation to stochastic noise falls below 1, the model should not be trusted to follow instructions in that environment. Instruction canonicalisation reduces sensitivity by 60–90% for two of three tasks under clean conditions. These findings, while specific to OpenVLA-7B, provide a methodology for evaluating compounded vulnerabilities and indicate that harsh-environment deployment requires degradation-type-specific evaluation.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Paper Acceptance: No

Submission Number: 14