Lights, Camera, Malfunction: When Illumination Robustness Leaves VLA Models Blind to Color

2026-07-16

AI Summary

This work addresses the vulnerability of Vision-Language-Action (VLA) models to real-world lighting perturbations and reveals that existing adversarial training methods, due to excessive augmentation, cause a collapse in color perception—reducing models to shape-only processors. To tackle this issue, the authors introduce FLARE, a physically grounded lighting attack framework, which for the first time exposes the color perception degradation induced by adversarial training. They further propose ChromaGuard, a chrominance-preserving adversarial training approach that enhances robustness to illumination variations while retaining critical color information. Experiments on a 6-degree-of-freedom robotic platform demonstrate that ChromaGuard achieves success rates of 97.5% and 92.5% on benign and attacked color-dependent tasks, respectively, substantially outperforming current baselines.

Abstract

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot manipulation; however, their transition to real-world environments reveals vulnerabilities to minor environmental perturbations. We propose FLARE, an optimized physical spotlight attack framework that exploits these vulnerabilities via targeted illuminations, dropping baseline task success rates to zero without any access to model internals. While adversarial training is the standard countermeasure, we identify a critical and previously underestimated defensive pitfall: naive data augmentations incorrectly condition VLA models to discard color as noise, collapsing their visual perception into a purely shape-biased processor. We expose this degradation through a diagnostic grayscale evaluation, in which the defended model maintains high success rates on grayscale inputs, while its success rate on benign, color-dependent real-world tasks drops to at most 47.5%, well below the undefended baseline. To address this, we propose ChromaGuard, a chroma-preserving adversarial training method. On a physical 6-DoF robotic platform, we demonstrate that ChromaGuard achieves 97.5% and 92.5% success rates in benign and attacked color-dependent tasks, respectively.

Problem

• Vision-Language-Action models
• illumination robustness
• color perception
• adversarial training
• environmental perturbations

Innovation

• adversarial training
• illumination robustness
• color perception
• Vision-Language-Action models
• physical attack