House3D is a large-scale virtual 3D simulation environment designed to support research in embodied AI, reinforcement learning, and vision-language navigation. It provides more than 45,000 richly annotated indoor scenes sourced from the SUNCG dataset, covering diverse architectural layouts such as studios, multi-floor homes, and spaces with detailed furnishings and room types. Each environment includes fully labeled 3D objects, allowing agents to perceive and interact with their surroundings through multiple sensory modalities including RGB images, depth maps, semantic segmentation masks, and top-down maps. The simulator is optimized for high-performance rendering, achieving thousands of frames per second to enable efficient large-scale training of RL agents. House3D has served as the foundation for several influential research projects such as RoomNav (for concept-based navigation) and Embodied Question Answering (EQA).
Features
- Over 45,000 annotated indoor 3D scenes with diverse layouts and object types
- Multiple observation modalities: RGB, depth, segmentation masks, and 2D maps
- High-speed rendering engine supporting thousands of frames per second
- Fully labeled 3D objects for scene understanding and object detection tasks
- Ideal for large-scale reinforcement learning and embodied AI experiments
- Supports key research tasks like RoomNav and Embodied Question Answering
Project Samples
