Closing the Train-Test Gap in World Models for Gradient-Based Planning
The paper addresses a mismatch between training world models and using them for planning, which causes poor performance. It proposes Online World Modeling (OWM) to correct predictions during planning and Adversarial World Modeling (AWM) to smooth the optimization landscape. These methods significantly improve gradient-based planning, allowing it to match or beat sampling-based planners while reducing computation time.