SceneBloom: AI-generated 3D scenes beat real scans for embodied training with 1.7x navigation gain
Lianhuiq · x · 2026-10-29
SceneBloom is an agentic framework generating realistic, diverse, simulation-ready 3D indoor scenes for embodied learning, asking whether generated scenes can train agents better than real data.
Method: a hierarchical sampler uses diffusion priors to generate furniture arrangements within functional zones; a compiler then uses extracted relations to guide geometric fitting and local pose correction, turning visual outputs into physically valid 3D scenes.
Results: across five room types it achieves the highest visual quality and physical stability, is the only method with no detected mesh collisions or out-of-bounds objects, and produces layouts closest to real rooms. Navigation policies trained in its scenes achieve 1.7x the navigation gain on held-out real scans versus an equal number of real training scans. Paper and code are public.
More from Embodied
- Robotics team hosts Sim2Real playoff: win in sim, control a real robot — cixliv · 2026-10-29
- NVIDIA's ProtoMotions lets you race the same policy that drives a real Unitree G1 — xbpeng4 · 2026-10-29
- Confetti: a YC-backed platform that turns plain English into shipped hardware parts — ycombinator · 2026-10-29
- Waymo doors won't open or close, support never calls — rider may switch back to humans — henloitsjoyce · 2026-10-29
- SuperMimic demo: vision-conditioned pi+ traverses stepped pit grid in sim — ChongZzZhang · 2026-10-29
- Machenta shows robots learning new tasks from a single demonstration — Ed__Johns · 2026-10-29