Obstacle avoidance using stereo vision and deep reinforcement learning in an animal-like robot

Obstacle avoidance is a fundamental behavior required to achieve safety and stability in both animals and robots. Many animals perceive and safely navigate their environment using two eyes with overlapping visual fields, allowing the use of stereopsis to compute distances to surfaces and to support collision avoidance. In this paper we develop an obstacle avoidance behavior for the biomimetic robot MiRo that combines stereo vision with deep reinforcement learning. We further show that avoidance strategies, learned for a simulated robot and environment, can be effectively transferred to a physical robot.