AdaReP: Plug-and-Play Acceleration for World Model Predictive Control using Adaptive Re-Planning

ICLR 2026 Conference Submission12974 Authors

Keywords: model predicative control, world model, robot manipulation

Abstract: We investigate the integration of model predictive control (MPC) with world models for robotic control tasks. Existing MPC solvers often replan at every step or after very few steps, primarily to mitigate the accumulation of world model prediction errors. However, such frequent replanning incurs substantial computational costs -- especially when using large, complex world models. In this work, we theoretically characterize the fundamental trade-off between computational efficiency and control performance in MPC. Our analysis reveals how replanning frequency, model prediction error, and local dynamics sensitivity jointly influence MPC performance, as captured by regret bounds. Based on the analysis, we propose AdaRep, a novel adaptive replanning mechanism for MPC that dynamically modulates the replanning frequency based on online estimates of world model prediction error and local dynamics sensitivity. AdaRep is training-free, plug-and-play, and compatible with various world models and MPC solvers. Experiments on the VP2 simulation benchmark across diverse tasks, as well as real-world robotic tasks including door opening and T-block pushing, show that AdaRep achieves substantial reductions in computation, over 80–90% in the real-world settings while maintaining or improving task success rates.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 12974