Human-Vehicle Shared Steering Control for Obstacle Avoidance: A Reference-Free Approach With Reinforcement Learning

Although artificial intelligence has made tremendous progress recently, there remain various technical issues and ethical problems before autonomous vehicles can be available to the general public. In the interim, collaboration between automated systems and human drivers becomes a promising solution, where the merits of machine intelligence and human intelligence are blended in a complementary way. To this end, the paper proposes a reference-free human-vehicle shared control framework based on reinforcement learning. Firstly, a personalized human-like driver agent is derived from highway driving data by means of generative adversarial imitation learning integrated with Gaussian mixture model. The driver model is responsible for real-time interaction with the reinforcement learning agent to relieve the burden of human operators in the course of training. Then, a copilot agent learns the policies to cooperatively control the vehicle steering based on three distinct implementations for the search of the best coordination strategy. Heuristic reward functions are designed to guide the agent policy optimization for multi-objective equilibrium between driver synchronization against intervention. To verify the control performance of the proposed shared driving system, simulation experiments with driver models and human-in-the-loop tests with real-life participants are conducted in the end of this paper. The results demonstrate that the shared steering control method can effectively follow human intentions, facilitate driving goals, improve road safety and reduce driver’s workload simultaneously in the challenging dynamic obstacle avoidance scenarios.