RedVLA: Physical Safety Red Teaming for VLA Models

The researchers propose RedVLA, which they state is the first red-teaming framework targeting the physical safety of vision-language-action (VLA) models, used to proactively discover unsafe behaviors that may cause irreversible physical harm before deployment. The framework has two stages: risk scenario synthesis identifies key interaction regions from benign trajectories and places risk factors so that they become entangled with the model's execution flow to induce the target unsafe behavior; risk amplification then iteratively adjusts the states of the risk factors through trajectory-feature-guided gradient-free optimization, stably triggering them across heterogeneous models. Experiments on six representative VLA models show that RedVLA can induce diverse unsafe behaviors, with an attack success rate of up to 95.5% within 10 rounds of optimization. The authors also propose SimpleVLA-Guard, a lightweight safety guard built with data generated by RedVLA; the data, assets, and code have been open-sourced on the project website redvla.github.io.