RoboSafe: Protecting Embodied Agents with Executable Safety Logic

RoboSafe is a runtime safety protection framework that protects VLM-driven embodied agents through executable safety logic. It runs on hybrid long-term and short-term safety memory, performing backward reflective reasoning (tracing recent trajectories and triggering replanning upon detecting a violation) and forward predictive reasoning (anticipating risks by combining long-term memory with multimodal observations). Experiments show that it reduces dangerous behaviors by 36.8% compared with leading baselines, with task performance close to the original level, and its practicality is validated on a physical robotic arm.