SceneBloom: agentic diffusion-generated 3D scenes give 1.7x navigation gain over real scans
Lianhuiq · x · 2026-11-01
NYU and UCSD researchers introduce SceneBloom, an agentic framework that generates realistic, simulation-ready 3D indoor scenes for embodied learning: diffusion models propose functional zones (seating, dining, storage), which are compiled into a scene; resampling the same prompt yields diverse layouts.
Key results:
- 100% physically valid scenes (no collisions, no out-of-bounds objects) vs 22% for Holodeck/SAGE and 18% for SceneWeaver;
- Layouts closest to real rooms among compared generators;
- Object navigation policy trained on its scenes reaches 35.3% success on held-out real scans, beating the 33.0% of equal-sized real-scan training — about 1.7x the navigation gain.
Takeaway: generated scenes can train embodied agents better than real-world data. Paper, code and project page are public.
More from Embodied
- Human2Sim demo adds deformable objects for robot sim-to-real training — danfei_xu · 2026-11-01
- Fauna Robotics robot dances into Halloween with new demo — lucacarlone1 · 2026-11-01
- Corebrum Physical AI OS adds ARM64 support: ROS plus AI agents on Raspberry Pi — chrismatthieu · 2026-11-01
- Hacker Turns the WebAI Badge Into a Chess Puzzle Device — jason_mayes · 2026-11-01
- Possessed Reachy robot becomes a multiplayer AR boss you can fight with a friend — stspanho · 2026-11-01
- DeepMind team builds dancing 12-DoF robot from a single English prompt, no CAD — GoogleDeepMind · 2026-11-01