Quadrotor Path Following and Reactive Obstacle Avoidance with Deep Reinforcement Learning

Abstract A deep reinforcement learning approach for solving the quadrotor path following and obstacle avoidance problem is proposed in this paper. The solved with two agents: one task another task. novel structure proposed, where action computed by agent becomes state of agent. Compared to traditional approaches, method allows interpret training process outcomes, faster can be safely trained on real quadrotor. Both agents implement Deep Deterministic Policy Gradient algorithm. was developed a previous work. uses information provided low-cost LIDAR detect obstacles around vehicle. Since has narrow field-of-view, an providing memory previously seen developed. detailed description defining vector, reward function given. are programmed python/tensorflow tested RotorS/gazebo platform. Simulations results prove validity approach.