SceneBloom generated robot training rooms with no mesh collisions, and policies trained there beat real scans
AspireAI community · arXiv (Yuhan Liu, Lianhui Qin, Shenji Wan)
Generalizable embodied agents need indoor environments that are realistic, varied and simulation-ready, but each route into synthetic scenes trades something away. Language-guided methods handle semantic planning and object-level reasoning yet struggle with precise spatial grounding, while visual generative models produce plausible arrangements whose output still needs geometric reasoning to become physically valid 3D scenes. SceneBloom combines the two: a hierarchical sampler uses diffusion priors to place furniture inside functional zones, and a compiler uses relations extracted from each proposal to drive geometric fitting and local pose correction. Across five room types it attained the highest visual quality and physical stability among the compared methods, and was the only one with no detected mesh collisions and no out-of-bounds objects. Its layouts also sat closer to the distribution of real rooms than every compared baseline. The result that matters most is downstream: object navigation policies trained in these generated scenes achieved a higher success rate on held-out real scans than policies trained on an equal number of real training scans. The honest caveat: the comparisons are against the methods the authors selected, and the real-scan transfer result is specific to the object navigation task and simulator configuration they evaluated rather than a general claim about sim-to-real.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.