Physical Experience Builds Continual Spatial World Models

This research investigates how embodied agents can build continually useful spatial world models by integrating repeated physical experience. Using a physics-grounded dynamic maze, the study distinguishes between persistent knowledge and adaptable understanding of environmental changes. Key findings show that 'Cross-episode Spatial Change Memory' learns relations among transformations, while state-action-outcome data refines motion predictions. The results indicate that predictive memory adjusts behavior when previous decisions become invalid but maintains stability when they remain correct. However, the advantage vanishes if past changes offer no information about future ones, and correct topology alone does not guarantee accurate action prediction without accumulated physical experience to reduce mismatches within modeled dynamics.