Research on Decision-Making Method of Unmanned Tractor-Trailer Based on T-DDPG

Abstract

Deep reinforcement learning has an excellent performance in decision-making and is widely used in areas such as autonomous driving. To improve the ability of decision-making of tractor-trailers, this paper presented an unmanned tractor-trailer decision-making model based on Deep Deterministic Policy Gradient (DDPG) algorithm. The DDPG algorithm has the problems of slow convergence in the early stage, poor stability, and easily falling into the local minimum value. Combined with the kinematics characteristics of the tractor-trailer, an improved DDPG algorithm is proposed. By adding improved artificial potential fields to the DDPG algorithm, the local minimums can be avoided. Compression of state set and action set of agent improves training speed. The stability of the algorithm is improved by adding a penalty item for a large angle between tractor and trailer and a penalty term for deviating from the desired trajectory to improve algorithm stability. Experimental results showed that the improved model improves learning efficiency, security, and average velocity of the tractor-trailer while ensuring the effectiveness of decision-making.