WALL-WM Replaces Fixed-Chunk VLA Training with Event-Grounded Action Units

WALL-WM reframes video-action learning by replacing fixed-length action chunk prediction with semantically coherent event boundaries as the atomic learning unit, addressing the granularity mismatch between language goals, visual dynamics, and control-level actions. The approach uses event-grounded Vision-Language-Action pretraining rather than initializing from standard multimodal foundation models.