Kairos-HomeWorld – Daxiao Robotics' Whole-Home 3D Interactive World Model

1. What Is Kairos-HomeWorld

Kairos-HomeWorld is the world's first whole-home 3D interactive world model, jointly released by Daxiao Robotics with The Chinese University of Hong Kong and Shenzhen Hetao Institute. Through an innovative four-stage hierarchical generation architecture, it delivers end-to-end generation from text descriptions to structurally coherent, physically compliant, object-interactive complete residential 3D scenes. It simultaneously open-sources a large dataset of 300,000 real Chinese floor plans and 5,000 whole-home simulation scenes, providing an embodied AI training foundation aligned with local living environments and addressing the mismatch caused by reliance on Western datasets.

Image source: Official article

Technical positioning and domain: It sits at the intersection of embodied AI and 3D vision, focused on large-scale, interactive, physically realistic indoor scene generation. It is not merely a 3D asset generator but a world-model platform for robot simulation training and smart-home prevalidation, emphasizing global consistency, physical compliance, and object-level interactivity.

Development background: Led by Daxiao Robotics with top research teams from CUHK and Hetao Institute. The motivation is embodied AI's "data gap"—Western datasets poorly match Chinese home environments, hurting robot training. The team leverages robotics, computer vision, and NLP to build a high-fidelity, low-cost, scalable virtual training environment designed for Chinese households.

Core value: Kairos-HomeWorld addresses three embodied AI pain points: lack of large-scale, high-quality Chinese home scene data; existing methods failing to guarantee whole-home global coherence and physical plausibility; and static "shell" 3D scenes without physically interactive objects. End-to-end interactive scene generation greatly lowers sim-to-real cost and accelerates home service robot commercialization.

Technical characteristics: The core is a four-stage hierarchical architecture decoupling whole-home generation into floor-plan structured learning, top-down global initialization, VLM physical compliance validation, and surface-centered object placement. It also introduces K-D tree structured representation and surface-centered object placement algorithms, leading in topology, geometric precision, and physical interaction.

2. Key Features

• End-to-end whole-home 3D generation: Natural language instructions trigger automatic pipeline from floor-plan understanding to complete 3D scene construction, covering 30 m² studios to 200 m²+ large flats with coherent structure, physical plausibility, and functional completeness—solving single-room generation and weak global consistency in traditional methods.

• Object-level full interactivity: Each scene averages 15+ objects with full physical properties (density, hinges, materials) importable into engines such as Physx-Omni for grasping, moving, stacking, and opening cabinets/doors, supporting long-horizon household tasks.

• Open Chinese home dataset: 300,000 structured real residential floor plans and 5,000 whole-home simulation scenes—the largest open dataset of its kind globally—capturing north-south ventilation, enclosed kitchens, separate balconies, wet/dry separation, and other Chinese housing traits as a reliable national data foundation.

• Physical compliance and closed-loop validation: A fine-tuned VLM module recursively validates layouts, fixing conflicts like "sofa blocking door" or "objects clipping through walls," keeping furniture collision rates at industry-leading levels under real physical rules.

• Full robot simulation training support: Beyond scenes, it supports long-horizon cross-room navigation and multi-room tidying in simulation—from generation through planning to policy execution in a closed training platform, shortening sim-to-real deployment.

• Four-stage hierarchical architecture: The technical engine decouples generation into K-D tree structured learning, top-down initialization and first-person roaming, VLM physical compliance validation, and surface-centered object placement—improving quality and efficiency with independently optimizable stages and strong extensibility.

3. How to Use

1. Resources and prerequisites: Download datasets, model weights, and docs from https://kairos-homeworld.github.io/. Recommended: NVIDIA A100 or equivalent server GPU with ≥40 GB VRAM. Software: CUDA 11.8+ PyTorch and Physx-Omni Python API.

2. Environment and deployment: Extract weights, pip install -r requirements.txt, then python deploy.py --config config.yaml. Key config: scene_type (e.g., apartment), floor_area (e.g., 120), num_rooms, etc.

3. Generation and instructions: After service start, send natural language via REST API or CLI, e.g., "Generate a 120 m² three-bedroom, two-living-room, north-south ventilated modern home." Four-stage generation runs unattended and outputs .usd or .glb in minutes.

4. Simulation and training: Import scenes into Physx-Omni, instantiate interactive objects with scene.instantiate(), connect robot via ROS2, and run long-horizon tasks like "bring the cup from kitchen to living room coffee table."

5. Sim-to-real and fine-tuning: Export policies with policy.export(), deploy to real robots, and fine-tune in real homes. Scenes align with real physical and layout logic, lowering migration cost.

4. Pros and Cons

5. Comparison with Similar Tools

Selection advice:

• Chinese home robot R&D: Kairos-HomeWorld is the top choice. Localized data and end-to-end generation best address the data gap for long-horizon Chinese household tasks.

• Physics fidelity and generality: For fine manipulation, fluids, soft bodies without strong regional needs, ThreeDWorld excels with top physics but manual scene building and lower automation.

• Fast start and standardized benchmarks: Habitat-Sim + AI2-THOR is mature for algorithm comparison with high scene quality and community, but limited scene count and no auto-generation. ProcTHOR scales procedural generation but lags Kairos-HomeWorld on quality and Chinese home support.

6. Editor's Take

Kairos-HomeWorld is a major breakthrough on embodied AI's core bottlenecks of data and scenes. The four-stage hierarchical architecture decouples complexity more wisely than black-box end-to-end models. VLM closed-loop physical validation ensures usability—a step many peers skip.

Practical value lies equally in the 300K floor plans + 5K simulation scenes dataset unmatched in scale and local fidelity, solving "no rice, no cooking" for Chinese embodied AI firms and cutting environment build cost. It is poised to become infrastructure for home service robot R&D in China.

The positioning is precise: embodied AI researchers, robot R&D teams, and smart-home designers. Academia gets a benchmark platform; industry gets an accelerator. With community growth and third-party tools, it could evolve from scene generator to embodied AI operating system.

Hardware barriers and scene detail richness remain, but strengths dominate.

Rationale: Top-tier architecture and dataset scale with pioneering localization; −0.5 for high hardware needs and early community, creating barriers for some SMBs and indie developers. Overall a milestone work worth deep study for embodied AI and home robotics.

7. Use Cases

• Home service robot training: Train vacuum path planning across rooms or tidying robots on long-horizon tasks like "return books from coffee table to shelf" in diverse Chinese home scenes with shorter sim-to-real cycles.

• Embodied AI R&D platform: Virtual training ground accelerating algorithm iteration without manual sim environment builds—focus on algorithm innovation at lower cost.

• Smart home prevalidation: Simulate furniture layout and device flows; e.g., place robot dock and air purifier in living room and simulate cleaning paths and obstacle avoidance to optimize space and smart-home plans.

• Academic research and benchmarks: Large-scale standardized datasets for indoor scene understanding, path planning, and manipulation with reproducible cross-model comparisons.

• Real estate and interior preview: Interactive 3D walkthroughs from real floor plans—open cabinets, move furniture—beyond static renders.

8. FAQ

Q: Is the Kairos-HomeWorld dataset free and open source?
A: Yes. Under open license, 300K Chinese floor plans, 5K whole-home simulation scenes, model weights, and docs are free on the site and GitHub for commercial and academic use.

Q: How long to generate a whole home? Hardware requirements?
A: Depends on complexity and hardware. A standard 120 m² three-bedroom flat on A100 typically takes 5–10 minutes. Recommend ≥40 GB VRAM; RTX 3090/4090 can run but may be slower or OOM on large scenes.

Q: Biggest advantage vs ProcTHOR, AI2-THOR?
A: Two points: localized data for Chinese homes vs Western mismatch, and whole-home end-to-end generation with coherent structure, physical compliance, and built-in interactive objects—not single rooms only.

Q: Are all objects interactive? What operations?
A: Average 15+ objects with physics (collision, gravity, friction): grasp, move, stack, rotate, open/close cabinets and drawers. Import into Physx-Omni-compatible engines.

Q: Can indie developers without expensive servers use it?
A: Yes. Download datasets for local algorithm work. For generation, use cloud GPU on demand or start with pre-generated sample scenes without full local deployment.

Q: Chinese natural language support?
A: Strong support for Chinese home descriptions, e.g., "north-south ventilated apartment with enclosed kitchen," thanks to Chinese corpus and localized training data.

9. Project Links

• Official site: https://kairos-homeworld.github.io/
• GitHub: https://github.com/Kairos-HomeWorld/HomeWorld
• arXiv paper: https://arxiv.org/pdf/2606.06390