RFW-W: Scaling Physical Asset Universes for Generative Embodied Simulation

Broad object coverage and executable physical representations are essential for scaling embodied simulation. Generative 3D models expand visual diversity, but scene construction and manipulation also require physical and interaction annotations. We present RFW-W, which combines RFW-R, a large-scale rigid-object dataset, with a common physical-asset interface and task-conditioned scene composition. RFW-R contains 524,611 retained assets across 1,212 categories, with collision geometry, composed physical properties, and interaction annotations exposed through standardized simulation interfaces. For each task, the system retrieves semantically relevant, physically compatible assets and assembles them around the specified object relations. For scene construction on 16 shared room layouts, RFW-R supplies and places 100% of the 265 requested objects, compared with 97.4% for EmbodiedGen v2 and 90.6% for ManiTwin. Across five RoboTwin tasks with 100 demonstrations per task and training source, RFW-R-trained policies achieve 71.2% clean and 41.8% randomized macro success, exceeding the strongest baseline by 15.8 and 10.6 percentage points, respectively. These results demonstrate the utility of RFW-R for scene construction and simulated policy learning.