AdaReP:Adaptive Re-Planning under Model Mismatch for Neural World-Model Predictive Control

The paper introduces AdaReP, a training-free wrapper designed for neural world-model predictive control (MPC) that adaptively adjusts replanning tolerance based on real-time deviations from cached rollouts and local dynamics sensitivity. This approach significantly reduces computational overhead associated with replanning, achieving over 80% fewer queries in a physical robot study while maintaining comparable task performance across various planning scenarios, including image-space and latent-space control. AdaReP's ability to optimize replanning without altering the underlying world model or planner is crucial for practitioners seeking efficient implementations of MPC in dynamic environments.