The paper introduces State-Aware Fallback Imitation (SAFI) to address the failure of Vision-Language-Action (VLA) models to safely halt prohibited behaviors mid-execution. Evaluation on OpenVLA-OFT reveals that policies suppressing tasks from initial states still complete them when taking over at intermediate states, and simple non-completion is insufficient for safety as robots may drop held objects. SAFI operates in two stages: fallback target construction assigns stationary arm commands and consistent gripper targets (holding) throughout prohibited-task demonstrations, and joint policy adaptation learns these fallback targets alongside original actions for retained tasks. This teaches an explicit safe response at intermediate states without degrading other skills. Experiments across LIBERO and BadLIBERO benchmarks show SAFI reduces aggregate target success after takeover to 1.3% with zero physical incidents or action saturation, while preserving high success rates on retained tasks.

Read original