EA-WM: Event-Aware World Models with Task-Specification Grounding for Long-Horizon Manipulation

The article introduces EA-WM, an event-aware world model framework designed to enhance long-horizon manipulation tasks by integrating task-specification-grounded event prediction and verification with pretrained visual-feature dynamics. EA-WM evaluates candidate futures in visual-feature space, transforming them into structured event states and scoring them based on task progress, semantic consistency, physical feasibility, and uncertainty. This approach improves the interpretability and alignment of world models with task requirements, demonstrating effectiveness across various manipulation scenarios, including navigation and deformable-object tasks, which is significant for practitioners seeking to enhance the reliability of robotic planning and execution.