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

AdaReP is a training-free algorithm that optimizes neural world-model predictive control by dynamically deciding when to replan versus reusing cached plans. By analyzing prediction mismatch propagation through local dynamics, the method achieves over 80% reduction in computational queries while maintaining task performance across simulated and real robotic manipulation tasks.

AdaReP addresses a fundamental efficiency challenge in neural world-model based control systems. Current approaches either replan at every step—incurring heavy computational costs—or cache plans and risk accumulating prediction errors. This research quantifies the trade-off through a perturbation-based dynamic-regret framework, providing theoretical grounding for when plan reuse becomes problematic. The framework reveals that stale-plan penalties depend on three factors: reuse tolerance, accumulated mismatch since last replanning, and local dynamics sensitivity. Rather than modifying underlying models or planners, AdaReP wraps existing systems as a lightweight decision layer that monitors deviation from cached rollouts and estimates local sensitivity in real time. This design choice matters significantly for practitioners—it enables adoption without retraining expensive neural world models. Across three distinct domains—image-space planning, latent-space control, and physical robot manipulation—AdaReP achieves substantial computational savings. The 50-trial physical robot study showing 80% fewer planner queries demonstrates the method's practical viability beyond simulation. For robotics and autonomous systems, reducing computational overhead directly translates to faster decision-making, lower energy consumption, and deployment feasibility on resource-constrained platforms. The work bridges theoretical analysis and engineering pragmatism, offering concrete efficiency gains without sacrificing task performance. Future research likely focuses on scaling this approach to longer-horizon tasks, more complex dynamics, and multi-agent scenarios where computational efficiency becomes even more critical.

• →AdaReP reduces neural world-model planner queries by over 80% on physical robots while maintaining performance parity.
• →The method adapts replanning tolerance online using deviation monitoring and sensitivity estimation without retraining models.
• →A perturbation-based framework quantifies how prediction mismatch, accumulated error, and local dynamics affect stale-plan penalties.
• →As a training-free wrapper, AdaReP integrates with existing planners and learned world models across different representations.
• →Efficiency gains enable practical deployment of neural predictive control on resource-constrained robotic platforms.