An End-to-End Deep Reinforcement Learning Model Based on Proximal Policy Optimization Algorithm for Autonomous Driving of Off-Road Vehicle

Most conventional unmanned vehicle control algorithms require human adjustment of parameters and design of precise rules, thus failing to adapt quickly to multiple situations when facing complex environments in the wild. To address these problems, this paper adopts an end-to-end deep reinforcement learning model based on proximal policy optimization algorithm to control the steering, speed and braking of an unmanned vehicle, allowing it to autonomously learn motion control strategies from perceptionmap in un-known environments. A novel environment simulator which contains variable passable areas and obstacles is also proposed to support agents to achieve target reward. The proposed agent model has been proved to receive the highest reward over SAC and has the ability to overcome the complexity of the wild environment generated by the simulator.