Leveraging Online Semantic Point Fusion for 3D-Aware Object Goal Navigation

Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, He Wang

Keywords: Reinforcement Learning, Robot, Navigation

Abstract: Object goal navigation in unseen environments is a fundamental task for building intelligent embodied agents. Existing works tackle this problem with modular or end-to-end learning-based methods, which implicitly learn from 2D maps, sparse scene graphs or video sequences, ignoring the established fact that objects lie in 3D. Hence, in this work, we propose a dedicated 3D-aware online semantic point fusion algorithm that online aggregates 3D points along with their semantic predictions from RGB-D observations to form a high-efficient 3D point-based sparse map, which further enables us to check spatial semantic consistency. To leverage the 3D information for navigation while remaining sample efficient, we then propose a two-stage reinforcement learning framework that decomposes the object goal navigation into two complementary sub-tasks, namely exploration and verification, each learning in a different discrete action space. Thanks to the highly accurate semantic understanding and robust goal verification, our framework achieves the best performance among all modular-based methods on the Matterport3D and Gibson datasets. Furthermore, compared to mainstream RL-based works, our method requires (5-28x) less computational cost for training. We will release the source code upon acceptance.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)

TL;DR: We propose a two-stage reinforcement learning framework that is powered by an online semantic point fusion algorithm.