End-To-End Deep Reinforcement Learning for First-Person Pedestrian Visual Navigation in Urban Environments

We solve a visual navigation problem in an urban setting via deep reinforcement learning in an end-to-end manner. A major challenge of a first-person visual navigation problem lies in severe partial observability and sparse positive experiences of reaching the goal. To address partial observability, we propose a novel 3D-temporal convolutional network to encode sequential historical visual observations, its effectiveness is verified by comparing to a commonly-used frame-stacking approach. For sparse positive samples, we propose an improved automatic curriculum learning algorithm NavACL+, which proposes meaningful curricula starting from easy tasks and gradually generalizes to challenging ones. NavACL+ is shown to facilitate the learning process, greatly improve the task success rate on difficult tasks by at least 40% and offer enhanced generalization to different initial poses compared to training from a fixed initial pose and the original NavACL algorithm.