Autonomous Driving for Natural Paths Using an Improved Deep Reinforcement Learning Algorithm

The purpose of this article is aimed to solve the problem associated with autonomous driving on the natural paths of planets. The contribution of this work is to propose an improved deep deterministic policy gradient (DDPG) framework for the autonomous driving on natural roads requires handling uneven surface of different throttle and braking reaction speeds. Our new finding is to design an adapted DDPG algorithm by double critic and excellent experience replay as DCEER-DDPG to reduce the overestimation of state action values. In addition, we created a virtual reality environment with TORCS simulator for fair evaluation. In the experiments, the proposed DCEER-DDPG has a better performance than previous algorithms, which can improve the utilization of driving experience on a natural path and increase the learning efficiency of the strategy. For the future applications, the proposed DCEER-DDPG is used not only on Earth, but also in lunar exploration.