House3D is a rich environment containing thousands of human-designed 3D scenes of visually realistic houses with fully labeled 3D objects, textures, and scene layouts. These virtual environments can be used to support novel research in deep reinforcement learning.
House3D contains over 45,000 human-designed 3D scenes that have been extracted from the SUNCG dataset. Scenes include single-room studios to multi-story houses, divided into room types.
All scenes in House3D have been annotated down to the level of individual objects, allowing agents to make observations of multiple modalities such as depth and top-down 2D views for deep reinforcement learning research.
Set up the SUNCG dataset following instructions in the SUNCGToolbox and organize the dataset.
SUNCG/ house 00065ecbdd7300d35ef4328ffe871505/ house.json house.mtl house.obj ... texture/ *.jpg
Compile the renderer and confirm it works
See tests/test-rendering.py for usage of the rendering API. See tests/test-env.py for an example use of the environment API.