Kairos-HomeWorld

Kairos-HomeWorld is a unified hierarchical framework for whole-home indoor scene generation, developed by ACE Robotics in collaboration with CUHK MMLab. It decomposes indoor scene synthesis into four controllable stages: floorplan generation, furniture layout distribution, VLM-based recursive refinement, and manipulable object placement. The project provides two large-scale resources: 300,000 real Chinese residential floorplans (nearly 4x the size of RPLAN, 17x of ResPlan) and 5,000 fully furnished whole-home 3D scenes with manipulable objects. All scenes have full physical attributes (material, density, friction coefficient) and support embodied AI simulation with object-level interaction. The generation pipeline uses an LLM trained with K-D tree representation for floorplan generation, foundation image models for hierarchical furnishing from top-down and ego-centric views, and a finetuned VLM refiner that iteratively fixes collisions (reducing collision rate from 0.20 to 0.05) and layout violations. This addresses the critical gap of Chinese home environments in embodied AI training — most existing indoor scene datasets are based on Western home layouts (open kitchens, no balconies, no entryways). The generated scenes support robot training for navigation, multi-room tidying, and household tasks.