Vision-Based Mobile Robotics Obstacle Avoidance With Deep Reinforcement Learning

Obstacle avoidance is a fundamental and challenging problem for autonomous\nnavigation of mobile robots. In this paper, we consider the problem of obstacle\navoidance in simple 3D environments where the robot has to solely rely on a\nsingle monocular camera. In particular, we are interested in solving this\nproblem without relying on localization, mapping, or planning techniques. Most\nof the existing work consider obstacle avoidance as two separate problems,\nnamely obstacle detection, and control. Inspired by the recent advantages of\ndeep reinforcement learning in Atari games and understanding highly complex\nsituations in Go, we tackle the obstacle avoidance problem as a data-driven\nend-to-end deep learning approach. Our approach takes raw images as input and\ngenerates control commands as output. We show that discrete action spaces are\noutperforming continuous control commands in terms of expected average reward\nin maze-like environments. Furthermore, we show how to accelerate the learning\nand increase the robustness of the policy by incorporating predicted depth maps\nby a generative adversarial network.\n