World-model planners evaluate candidate trajectories by predicting their future consequences, implicitly relying on a plan-to-motion relationship learned during training. This relationship is central to their ability to select good plans.
At deployment, however, a different downstream controller may be used, causing a mismatch between what the planner expects and what the controller actually executes. This execution shift can degrade performance, as the planner's predictions no longer align with real-world outcomes.
The paper, as its title suggests, tackles this problem by adapting the frozen planner to the new controller. The provided abstract cuts off before describing the adaptation method, but the work highlights a practical challenge in deploying world-model planners across changing control systems.