Quadrotor Mapless Navigation in Static and Dynamic Environments based on Deep Reinforcement Learning

Abstract

In this paper, we propose a mapless autonomous navigation planner which plans a collision-free trajectory for quadrotor without any manual operations. Deep Reinforcement Learning (DRL) can optimize the policy by trial and error without knowing the prior information of the environment. The designed reward function has better convergence which compares to the benchmark method. The learned policy makes a real time collision free trajectory which can cope with the dynamic obstacles under different scenarios. The evaluation result shows that the trained model can be applied directly to the unknown environment without retraining the agent.