Building Generalizable Agents with a Realistic and Rich 3D Environment.

Towards bridging the gap between machine and human intelligence, it is of utmost importance to introduce environments that are visually realistic and rich in content. In such environments, one can evaluate and improve a crucial property of practical intelligent systems, namely generalization. In this work, we build House3D, a rich, extensible and efficient environment that contains 45,622 human-designed 3D scenes of houses, ranging from single-room studios to multi-storeyed houses, equipped with a diverse set of fully labeled 3D objects, textures and scene layouts, based on the SUNCG dataset (Song et al., 2017). With an emphasis on semantic-level generalization, we study the task of concept-driven navigation, RoomNav, using a subset of houses in House3D. In RoomNav, an agent navigates towards a target specified by a semantic concept. To succeed, the agent learns to comprehend the scene it lives in by developing perception, understand the concept by mapping it to the correct semantics, and navigate to the target by obeying the underlying physical rules. We train RL agents with both continuous and discrete action spaces and show their ability to generalize in new unseen environments. In particular, we observe that (1) training is substantially harder on large house sets but results in better generalization, (2) using semantic signals (e.g., segmentation mask) boosts the generalization performance, and (3) gated networks on semantic input signal lead to improved training performance and generalization. We hope House3D, including the analysis of the RoomNav task, serves as a building block towards designing practical intelligent systems and we wish it to be broadly adopted by the community.