Cooperative Landing on Mobile Platform for Multiple Unmanned Aerial Vehicles via Reinforcement Learning

Abstract

This paper proposes a multiple unmanned aerial vehicles (UAVs) cooperative landing algorithm based on deep reinforcement learning. First, to solve the partial observation problem, we propose the recurrent neural network to predict the moving platform trajectory. Afterwards, with the centralized multiagent framework, we present a parameter sharing method to realize multi-UAV cooperation. Finally, focusing on the sensor noise problem of the actual UAV flight, we propose a noise compensation recurrent proximal policy optimization (NC-RPPO) algorithm to extract images’ features to compensate for inertial measurement unit (IMU) and GPS errors. We utilize AirSim to construct a simulated 3D environment resembling an offshore oil development zone. In this setting, we evaluate the effectiveness of our proposed multi-UAV cooperative landing algorithm while considering the presence of sensor noise. Through experimental trials, we demonstrate that our NC-RPPO algorithm enables UAVs to accurately predict the trajectory of a mobile platform and successfully land on it cooperatively in real time. Notably, the experimental outcomes obtained through our image-assisted noise correction method closely align with those obtained from the ground truth experiment.