Energy Minimization in Overtaking for Autonomous Vehicles in a Bidirectional Environment

In this article, we formulate an overtaking problem in a bidirectional dynamic highway environment. A Deep Deterministic Policy Gradient (DDPG)-based reinforcement learning method is used to learn an optimal policy for the Ego car with a customized environment. Moreover, the classical optimal control method is applied to solve a similar optimal control problem with two time-varying constraints. Simulations are provided to verify the performance of DDPG. The optimal policy obtained by the classical optimal control method is then used as a compare benchmark for the learning-based method. Model predictive path integral control is finally employed to handle a more dynamic environment and possible different driving modes of cars.