Back on Track: Deployment-time Policy Improvement via World Model Planning

This research proposes an offline-trained monitoring-and-intervention framework that improves visuomotor policies during deployment using black-box query access. The method employs a policy-conditioned liveness value and a compact latent world model to detect task progress loss and generate corrective actions in real-time, achieving significant improvements in task success rates on LIBERO-Goal and a Franka Panda manipulator.