Image-Goal Navigation via Keypoint-Based Reinforcement Learning

In this paper, we tackle the problem of image-goal navigation which is a crucial robot navigation task but also a hard problem especially when there exist obstacles and limitations on field-of-view (FoV) of the camera. Conventional visual servoing approaches require depth information and camera parameters, and are susceptible to FoV loss, while previous learning-based approaches depend on the unrealistic dense reward function to train the agent with reinforcement learning. To this end, we propose a novel reinforcement learning-based approach which simultaneously utilizes self-supervised local features and global features from an observed image and a target image. The proposed method, KeypointRL, exploits keypoint matching information and generates a self-supervised reward signal which allows the agent to be easily transferred to unseen environments. The proposed model is trained on the subset of image-goal dataset in the photo-realistic Gibson dataset together with Habitat simulator, and shown to outperform baseline algorithms and generalize better.