Vision-language-action (VLA) models let robots translate language commands into physical actions, but they are vulnerable to visual attacks that can redirect behavior. Earlier attack methods optimized perturbations against a fixed target, assuming the robot would follow a predictable trajectory. The new paper argues that this assumption is flawed: a perturbation changes early actions, which in turn changes the physical states the robot reaches later, so the original supervision no longer matches what is needed at those states.

To address this, the authors propose TRACO, a closed-loop optimization framework. Its Execution-Driven Supervision rebuilds supervision for the states actually reached under attack, guiding subsequent updates. A second component, Progress-Guided Selection, focuses updates on stages where progress has stalled while preserving the earlier behavior needed to reach that point. This prevents the attack from undoing its own successes.

In experiments across multiple VLA models and simulation environments, TRACO improved task-redirection success compared with existing targeted attacks. The authors highlight the security implications: closed-loop attack optimization can make visual perturbations more potent, suggesting that defenses for VLA-based robots need to account for adaptive, state-aware attacks.