Deep reinforcement learning conducts learning while automatically collecting data through iterative trials. It has attracted attention in recent years as a method for seeking optimal action policy and its applications have been increasing. However, in order to apply deep reinforcement learning to autonomous vehicle products that move under automatic control while satisfying multiple objectives, such as moving to a target position, avoiding obstacles, etc., there is the problem that the learning conditions must be appropriately adjusted before starting the learning. Digital innovation headquarter has taken on this challenge as a solution for our mission of smarter product and devised an advanced method to realize the easier application of deep reinforcement learning to autonomous vehicles and the more efficient improvement of the product performance, such as shortening of obstacle avoidance maneuvers. The effectiveness of avoiding obstacles efficiently under multiple obstacle conditions is confirmed by simulation and this report also presents the verification results.
