SceneBloom generates collision-free indoor layouts for embodied agent training
Original title: SceneBloom: Agentic Indoor Environment Generation with Realistic and Diverse Layouts for Embodied Learning
Navigation policies trained on its scenes beat policies trained on the same number of real scans in held-out scans.
Researchers introduce SceneBloom, an agentic framework that samples furniture layouts with diffusion priors and compiles them into physically valid 3D scenes using extracted spatial relations. Across five room types, the paper says it had the highest visual quality and physical stability among compared methods, with no detected mesh collisions or out-of-bounds objects. Object-navigation policies trained in these scenes reached a higher success rate on held-out real scans than policies trained on an equal number of real scans. Code is on GitHub.
- A hierarchical sampler places furniture inside functional zones.
- A compiler uses proposal relations for geometric fitting and pose correction.
- Layouts were closer to real-room distributions than all compared baselines.
- The authors report it was the only method with no detected collisions or out-of-bounds objects.
From the original article
1 min read at arxiv.orgTraining generalizable embodied agents requires realistic, varied, and simulation-ready indoor environments. Language-guided methods support semantic planning and object-level reasoning, but precise spatial grounding remains difficult. Visual generative models offer rich priors over plausible arrangements, yet their outputs require geometric reasoning to become physically valid 3D scenes.
By Yuhan Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.