ActFovea: Runtime Safeguarding for VLA Policies via Spatiotemporal Visual-Action Consistency

ActFovea is a newly introduced safeguarding framework designed to identify and address runtime issues in vision-language-action (VLA) policies utilized in robotic manipulation. Despite their effectiveness, these policies can experience failures that disrupt the synchronization of visual inputs, robot states, and performed actions. Notably, ActFovea functions without the need for retraining or altering the existing VLA policy. It utilizes robot kinematics, proprioceptive states, and prior actions to create action-conditioned foveated regions that focus on relevant areas and anticipated motion paths, while filtering out irrelevant visual information. The framework assesses runtime threats by checking the consistency of visual motion and observation recency against geometric, proprioceptive, and action shifts. For disturbances that can be recovered, ActFovea generates specific candidate observations and only accepts a recovery once action consistency is confirmed. This framework is elaborated in a paper available on arXiv (arXiv:2607.29169) and was announced as a cross-type submission, addressing a vital need for enhancing the reliability of VLA policies in practical robotic scenarios.