Lights, Camera, Malfunction: When Illumination Robustness Leaves VLA Models Blind to Color
Lights, Camera, Malfunction: When Illumination Robustness Leaves VLA Models Blind to Color
This research paper introduces FLARE, a framework demonstrating how targeted illumination can disrupt Vision-Language-Action (VLA) models. The authors identify that standard adversarial training often leads models to ignore color as noise, degrading performance on color-dependent tasks. They propose ChromaGuard, a chroma-preserving adversarial training method tested on a physical 6-DoF robotic platform using SmolVLA, which achieves 95.0% success in benign tasks and 88.8% under attack, compared to a maximum of 48.8% for naively defended models.
The study identifies a specific vulnerability in Vision-Language-Action models where naive data augmentation causes them to discard color information, leading to significant performance drops on real-world tasks; the authors propose ChromaGuard, an adversarial training method that preserves chromatic cues to maintain high success rates under both benign and attacked conditions.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
88.8%
Reported trials
View original evidence
le inputs, while its success rate on benign, color-dependent real-world tasks drops to at most 48.8%, compared with 78.8% for the undefended SmolVLA baseline. To address this, we propose ChromaGuard, a chroma-preserving adversarial training method. On a physical 6-DoF robotic platform using SmolVLA, we demonstrate that ChromaGuard achieves 95.0% and 88.8% success rates in benign and attacked color-dependent tasks, re
Open source S4robot platform degrees of freedom
6
View original evidence
le inputs, while its success rate on benign, color-dependent real-world tasks drops to at most 48.8%, compared with 78.8% for the undefended SmolVLA baseline. To address this, we propose ChromaGuard, a chroma-preserving adversarial training method. On a physical 6-DoF robotic platform using SmolVLA, we demonstrate that ChromaGuard achieves 95.0% and 88.8% success rates in benign and attacked color-dependent tasks, re
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
le inputs, while its success rate on benign, color-dependent real-world tasks drops to at most 48.8%, compared with 78.8% for the undefended SmolVLA baseline. To address this, we propose ChromaGuard, a chroma-preserving adversarial training method. On a physical 6-DoF robotic platform using SmolVLA, we demonstrate that ChromaGuard achieves 95.0% and 88.8% success rates in benign and attacked color-dependent tasks, re
Open source S4
spectively. Demo videos are available at https://keio-csg.github.io/ChromaGuard/.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org