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

2026-06-22

AI Summary

This work addresses the trade-off between computational efficiency and control performance in model-based reinforcement learning when combining neural world models with model predictive control (MPC). Frequent replanning incurs high computational costs, while reusing cached plans suffers from model mismatch, degrading performance. To resolve this, the authors propose AdaReP—an adaptive replanning framework that requires no additional training. AdaReP dynamically adjusts replanning tolerance by online estimation of trajectory deviation and local dynamics sensitivity, balancing efficiency and performance without modifying the world model or planner. The study introduces a perturbation-based dynamic regret analysis to quantitatively relate replanning frequency, accumulated model mismatch, and local sensitivity, informing the design of an online adaptation mechanism. Experiments demonstrate that AdaReP significantly reduces computational overhead—cutting planning queries by over 80% in physical robot tasks—while maintaining competitive task performance across image-space planning, latent-space control, and real-world robotic domains.

Abstract

Neural world models coupled with model predictive control (MPC) replan at every environment step to bound accumulated prediction error, but this incurs substantial computational overhead. Reusing a cached plan reduces this overhead, yet its effectiveness depends on how prediction mismatch propagates through the local dynamics. We analyze this trade-off with a perturbation-based dynamic-regret framework and show that stale-plan penalties scale with the reuse tolerance, the accumulated mismatch since the last replanning step, and the local dynamics sensitivity. Based on this structure, we propose AdaReP, a training-free wrapper that adapts the replanning tolerance online using the current deviation from the cached rollout and a local sensitivity estimate, without modifying the learned world model or planner. Across image-space planning, latent-space control, and real-world robotic manipulation, AdaReP substantially reduces planner-side computation while maintaining comparable task performance, including over 80% fewer queries on a 50-trial physical robot study.

Problem

• model predictive control
• neural world models
• replanning
• model mismatch
• computational overhead

Innovation

• adaptive replanning
• model predictive control
• neural world models
• dynamic regret
• computation efficiency