Visually Robust, Safety-Blind: Vision-Language-Action Policies Ignore the Robot Next to Them

Vision-language-action (VLA) policies are evaluated alone on a table, yet deployments increasingly place several robots in one workspace. We add a second, non-contacting Franka arm B to LIBERO scenes and evaluate two VLAs of very different competence, SmolVLA (~70% success) and π0.5 (~98%). A benign moving neighbor leaves task success unchanged for both, but the victim comes within 3 cm of B in 44.5–46.5% of episodes and drives into B's volume in 24–27%; a physical-collision check confirms these intrusions are real contacts. A whole-arm control-barrier-function (CBF) filter on the victim removes near-misses but halves success (68→26% and 98.5→48%) because both robots stall in front of each other. A simple coordination rule in which B yields along an augmented-Lagrangian retreat restores 85% and 98% of that lost success with zero near-misses over 400 episodes. Optimized adversarial motions for B did not beat motion-matched benign motion on held-out seeds: the danger is not a clever adversary but an oblivious victim.