Robots trained in structured environments often struggle in open-world conditions, where clutter, occlusion, and changing viewpoints obscure the spatial relationships needed for action. A new preprint introduces AWAM, an Active World-Action Model that addresses this gap by actively planning where to observe and imagining what will be seen along the planned trajectory before acting.
The method predicts task-related virtual camera motion and uses it as a structured query into a geometry-conditioned world model. This produces a reusable spatial prior for manipulation, which is then compressed into a fixed memory and fused with live observations through cross-attention for diffusion-based action prediction.
In evaluations, AWAM achieved 75.9% on LIBERO-Occ, 62.4% on LIBERO-Plus, 76.5% on RLBench-OG Occlusion1, and 94.4% in real-robot trials. These results suggest improved robustness to the evaluated occlusions and distribution shifts, helping close part of the gap between structured manipulation and less controlled environments.