Robust Deep RL-Based Aerial Transportation of Suspended Loads

In this work, a robust policy, trained using deep Reinforcement Learning (RL), is presented for the aerial transportation of cable-suspended loads with simultaneous minimization of the swinging motion of the cable. More precisely, domain randomization is applied throughout the learning procedure in a simulation environment in order to develop a policy which is robust to varying model parameters, e.g., load mass and cable length, as well as system dynamics that differ from the ones encountered during the training. Based on our approach, the gap between simulation and real-world conditions is bridged and the successful transfer of the policy, trained exclusively in simulation, to a real Unmanned Aerial Vehicle (UAV) is attained. The performance of the learned policy is demonstrated through both simulation and real-world experiments in an outdoor environment.