Autonomous Navigation and Control of a Quadrotor Using Deep Reinforcement Learning

A deep reinforcement learning-based control framework has been proposed in this paper to achieve autonomous navigation and control of a quadrotor. Cascaded reinforcement learning agents form the control framework. First, a path following (PF) agent controls the quadrotor's tracking behavior by directly mapping environment states into motor commands. The second agent modifies the desired path to avoid any detected obstacles along the path. The obstacle avoidance (OA) agent achieves this task by adding an offset distance deflection to the tracking error before sending it to the path-following agent. Generalization of the obstacle avoidance behavior in three-dimensional space was achieved by the usage of frame transformation. The two agents were trained using the "Twin Delayed Deep Deterministic Policy Gradient" (TD3) algorithm, and the developed framework succeeded in avoiding multiple obstacles of different sizes and configurations in simulation.