World action models (WAMs) aim to give robots a unified representation of visual state changes and actions, but existing versions struggle with long tasks, coarse language supervision, and text-only conditioning. WorldScape Policy 2.0 addresses these issues by adding a reasoning-augmented memory system. A causal short-term visual memory feeds recent observations into the model as diffusion-transformer prefill, while a long short-term event memory organizes past vision-language outputs into global, local-active, and event-boundary representations for progress-aware retrieval.
The retrieved history is used to augment perception and to generate planning tokens, creating an implicit subgoal condition for autonomous planning. Semantic forcing transfers event-level instruction semantics into this latent planning pathway, allowing the model to switch between high-level autonomous planning and execution conditioned on fine-grained text, goal images, or video context.
To support this controllability, the authors built ManipEvent-5M, a pretraining dataset with nearly five million event segments, each paired with action trajectories, episode-level instructions, segment-level subtask captions, goal images, and video demonstrations. Experiments in simulation and on real robots report improvements in long-horizon planning, fine-grained instruction following, and in-context adaptation.