Automatic control of unmanned vehicles using DPSN-DDPG

The deep deterministic policy gradient algorithm (DDPG) is a classical deep reinforcement learning (DRL) method that is particularly suitable for solving continuous control problems and has been widely used in autonomous driving. However, the traditional DDPG algorithm suffers from inefficient and ineffective exploration. To solve this problem, an improved DDPG algorithm with dynamic parameter space noise (DPSN) is proposed. In this paper, a framework that incorporates a variational autoencoder (VAE) and hierarchical normalized neural network with DPSN is constructed to achieve automatic control of unmanned vehicles. The experimental results on the Carla platform illustrate the DPSN-DDPG algorithm achieves significant improvements in exploration capability and training efficiency. The unmanned vehicles controlled by DPSN-DDPG can overcome local optimums, achieving higher rewards and completion rates.