Mapless navigation for UAVs via reinforcement learning from demonstrations

This paper is concerned with the problems of mapless navigation for unmanned aerial vehicles in the scenarios with limited sensor accuracy and computing capability. A novel learning-based algorithm called soft actor-critic from demonstrations (SACfD) is proposed, integrating reinforcement learning with imitation learning. Specifically, the maximum entropy reinforcement learning framework is introduced to enhance the exploration capability of the algorithm, upon which the paper explores a way to sufficiently leverage demonstration data to significantly accelerate the convergence rate while improving policy performance reliably. Further, the proposed algorithm enables an implementation of mapless navigation for unmanned aerial vehicles and experimental results show that it outperforms the existing algorithms.