en.Wedoany.com Reported - On June 5, DaXiao Robotics, in collaboration with the Multimedia Laboratory (MMLab, CUHK) at The Chinese University of Hong Kong and the Hetao Shenzhen Institute, released its world model research achievement, Kairos-HomeWorld. This is the world's first unified framework for world models that achieves full-home generation and full interaction with individual objects.
Existing indoor scene generation technologies can only cover single rooms, lacking global consistency and operability. Kairos-HomeWorld breaks through this industry bottleneck. The model adopts a four-stage hierarchical generation architecture of "global structure, local details, closed-loop verification, and interaction enhancement," achieving for the first time end-to-end generation of complete residential 3D scenes from a single text prompt—scenes that are globally consistent in structure, fully physically compliant, and with interactive objects—thereby reconstructing the technical paradigm of indoor scene generation.
In terms of home robot training, Kairos-HomeWorld leverages the model to batch-generate diverse simulated Chinese home scenes and objects with inherent physical interaction properties. Robots can perform various household training tasks in the virtual environment, with nearly zero marginal cost for new scenes, eliminating real-world expenses such as site maintenance and furniture wear and tear. Additionally, this method is not limited by the total number of real-world housing units, offering advantages in training efficiency and large-scale expansion over field data collection approaches.
Kairos-HomeWorld has already been applied to the embodied intelligence daily training of DaXiao Robotics, supporting robots in completing full-process simulation training for complex long-horizon household tasks such as cross-room navigation and multi-room object organization, shortening the migration cycle from virtual simulation to real-world deployment.
The research team has simultaneously open-sourced the world's largest full-home 3D dataset, the first specifically designed for Chinese households. This dataset includes 300,000 structurally annotated real residential floor plans, 5,000 full-home simulation scenes with complete interior layouts and furniture arrangements, and 50,000 object assets supporting physical simulation and interactive operations, covering various typical Chinese home layouts and fully restoring the living characteristics of Chinese households.
Among these, the 300,000 real residential floor plans are all sourced from genuine listings in the Chinese market. After a multi-stage automated processing pipeline for vectorization and structural annotation, they contain full-dimensional information including door and window positions, room geometry, functional zoning, and connectivity relationships, making it the world's largest dataset of real residential floor plans. The full-home simulation scenes with furniture and objects generated based on this training each include a complete furniture layout. Using the Physx-Omni model, each scene automatically generates an average of over 15 interactive objects with physical properties (including density, articulation, manifold, etc.), all of which can be directly imported into the simulation engine for interactive training.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
