SceneBloom: Agentic Indoor Environment Generation with Realistic and Diverse Layouts for Embodied Learning
SceneBloom combines diffusion-generated furniture layouts with a compiler that fits them into valid 3D scenes. Its results suggest a way to create varied training environments for embodied navigation agents.
SceneBloom samples furniture arrangements within functional zones, then compiles them into 3D using extracted spatial relations. Across **five room types**, the authors report no detected mesh collisions or out-of-bounds objects, and better held-out real-scan navigation success than training on an equal number of real scans.
Builders working on embodied agents can consider separating scene proposals from geometric validation and correction. The paper’s sampler and compiler illustrate that split for generating simulation-ready environments.
SceneBloom samples furniture arrangements within functional zones, then compiles them into 3D using extracted spatial relations. Across **five room types**, the authors report no detected mesh collisions or out-of-bounds objects, and better held-out real-scan navigation success than training on an equal number of real scans. Builders working on embodied agents can consider separating scene proposals from geometric validation and correction. The paper’s sampler and compiler illustrate that split for generating simulation-ready environments. The comparison is limited to the methods and settings studied. The supplied material gives no numerical navigation results or details on how the approach transfers to other environments.
SceneBloom’s sampler-to-compiler split makes geometric validation and correction a distinct stage in embodied-scene generation. Its reported collision and navigation results are specific to the studied room types and comparisons; the supplied candidates offer a useful reminder that evaluation should target the task’s concrete failure modes, but do not establish transfer to other environments.