Vision-based Collision Avoidance for Mobile Robots through Sim-to-Real Transfer

Jin-Woo Lee, Keon-Woo Kim, Soohyun Shin, Seong-Woo Kim

Abstract: In mobile robots, collision avoidance has been generally performed using range sensors, but due to its limits, additional RGB image-based control is required. In this paper, we propose an end-to-end visuomotor training method for avoiding collision based on deep reinforcement learning. We created a compact and intensive Gazebo simulation environment for training a robot and tested it both in a real environment and the simulated virtual environment. We demonstrated that the robot is able to successfully explore the test environment without collision, as well as obtain training results that can be generalized in a compact and randomized environment. The result shows the potential for learning-based collision avoidance in outdoor environments, since we do not utilize depth information.