Gasgoo Munich- ACE ROBOTICS, in collaboration with the Multimedia Laboratory at The Chinese University of Hong Kong (CUHK) and the Shenzhen Hetao Academy, unveiled its latest world model research achievement, Kairos-HomeWorld, on June 5.
This marks the industry's first unified world model framework capable of generating entire homes with fully interactive individual objects. It shatters existing bottlenecks where indoor generation is limited to single rooms and lacks global consistency and operability. With a single click, the system can produce structurally coherent, physically plausible, and functionally complete 3D whole-house scenes. This creates a massive, high-fidelity interactive training ground for embodied AI and robotics, specifically tailored to Chinese household environments.
Image Source: ACE ROBOTICS
Specifically, Kairos-HomeWorld leverages a four-stage hierarchical generation architecture: global structure, local details, closed-loop verification, and interaction enhancement. This enables the first end-to-end generation of complete residential 3D scenes from a single text prompt—ensuring global structural consistency, full physical compliance, and object interactivity. The approach fundamentally reconstructs the technical paradigm for indoor scene generation.
For home robot training, Kairos-HomeWorld offers distinct advantages in cost and efficiency. By relying on the model to batch-generate diverse Chinese home simulations and objects with inherent physical properties, robots can master various household chores within virtual environments. The marginal cost for new scenarios is near zero, eliminating real-world expenses like site maintenance and furniture wear. Furthermore, it isn’t constrained by the limited supply of physical housing, offering superior scalability and efficiency compared to traditional field collection methods.
Kairos-HomeWorld is already in active use for ACE ROBOTICS's daily embodied intelligence training. It supports full-process simulation for complex, long-horizon household tasks—such as cross-room navigation and multi-room organization. This capability significantly shortens the transition cycle from virtual simulation to real-world deployment, while markedly lowering the barriers to entry for embodied intelligence R&D.
Image Source: ACE ROBOTICS
Simultaneously, the research team open-sourced the world's largest whole-house 3D dataset—and the first built specifically for Chinese households. The dataset includes 300,000 structured, annotated real residential floor plans; 5,000 simulated whole-house scenes with complete interior layouts and furniture; and 50,000 object assets supporting physical simulation and interaction. Covering typical housing layouts nationwide, it faithfully recreates the living characteristics of local Chinese homes, providing the core data and technical foundation for the domestic deployment of embodied AI.
Notably, all 300,000 floor plans are sourced from real listings in the Chinese market. Processed through a multi-stage automated pipeline, they have been vectorized and structurally annotated to include comprehensive data—door and window positions, room geometry, functional zones, and connectivity. This stands as the world's largest dataset of real residential floor plans.
Building on this foundation, the generated furnished whole-house simulations feature complete layouts for each unit. Leveraging the PhysX-Omni model, the system automatically generates an average of 15 or more interactive objects per unit—complete with physical properties like density, articulation, and manifold types. All assets support direct import into simulation engines for interactive training.
