A new technique lets AI-controlled robots decide for themselves when they can skip replanning, cutting compute costs without hurting performance.
AdaReP tackles a core inefficiency in neural world-model control: recalculating a full plan at every single step to keep prediction errors from piling up. That constant replanning is safe but computationally expensive, and simply caching a plan only works if you know when it has drifted too far from reality. The researchers built a framework analyzing how prediction mismatch propagates through local dynamics, then used it to create AdaReP, a training-free wrapper that adjusts a plan's shelf life based on real-time deviation from the cached rollout and a local sensitivity estimate. It requires no changes to the underlying world model or planner, and was tested on image-space planning, latent-space control, and real-world robotic manipulation.
Compute is the bottleneck holding back real-time robot control with learned world models, since replanning every step can be prohibitively slow on physical hardware. AdaReP's results suggest that bottleneck is partly self-inflicted: in a 50-trial physical robot study, it cut planner queries by more than 80 percent while matching task performance.
It is a tuning trick, not a new world model or planner, but if the gains hold up outside curated benchmarks, the fix is refreshingly boring: plan less, not smarter.