DRL-based Path Planner and its Application in Real Quadrotor with LIDAR

Abstract

<p>The distribution mismatching issue has been hindering the landing of deep reinforcement learning algorithms in the robot field for a long time. This paper proposes a novel DRL-based path planner and corresponding training method to realize the safe obstacle avoidance of real quadrotors. To achieve the goal, we design a randomized environment generation module to fit the reality-simulation error. Then the map information can be parameterized to make the test data statistically significant. In addition, an instruction filter is proposed to smooth the output of the policy network in the test phase. Its improvement in obstacle avoidance performance is demonstrated in the experiment section. Finally, real-time flight experiments are conducted to verify the effectiveness of our algorithm and prove that the learning-based path planner can solve practical problems in the robot field. Our framework has three advantages: (1) map parameterization, (2) low-cost planning, and (3) reality validation. The video and code are available: .https://github.com/Vinson-sheep/multi rotor avoidance rl.</p>