NavRep: Unsupervised Representations for Reinforcement Learning of Robot Navigation in Dynamic Human Environments

Robot navigation is a task where reinforcement learning approaches are still\nunable to compete with traditional path planning. State-of-the-art methods\ndiffer in small ways, and do not all provide reproducible, openly available\nimplementations. This makes comparing methods a challenge. Recent research has\nshown that unsupervised learning methods can scale impressively, and be\nleveraged to solve difficult problems. In this work, we design ways in which\nunsupervised learning can be used to assist reinforcement learning for robot\nnavigation. We train two end-to-end, and 18 unsupervised-learning-based\narchitectures, and compare them, along with existing approaches, in unseen test\ncases. We demonstrate our approach working on a real life robot. Our results\nshow that unsupervised learning methods are competitive with end-to-end\nmethods. We also highlight the importance of various components such as input\nrepresentation, predictive unsupervised learning, and latent features. We make\nall our models publicly available, as well as training and testing\nenvironments, and tools. This release also includes OpenAI-gym-compatible\nenvironments designed to emulate the training conditions described by other\npapers, with as much fidelity as possible. Our hope is that this helps in\nbringing together the field of RL for robot navigation, and allows meaningful\ncomparisons across state-of-the-art methods.\n