Path planning via reinforcement learning with closed-loop motion control and field tests

Performing evasive maneuvers with highly automated vehicles is a challenging task. The algorithm must fulfill safety constraints and complete the task while keeping the car in a controllable state. Furthermore, considering all aspects of vehicle dynamics, the path generation problem is numerically complex. Hence its classical solutions can hardly meet real-time requirements. On the other hand, single reinforcement learning based approaches only could handle this problem as a simple driving task and would not provide feasibility information on the whole task’s horizon. Therefore, this paper presents a hierarchical method for obstacle avoidance of an automated vehicle to overcome this issue, where the geometric path generation is provided by a single-step continuous Reinforcement Learning agent, while a model-predictive controller deals with lateral control to perform a double lane change maneuver. As the agent plays the optimization role in this architecture, it is trained in various scenarios to provide the necessary parameters for a geometric path generator in a one-step neural network output. During the training, the controller that follows the track evaluates the feasibility of the generated path whose performance metrics provide feedback to the agent so it can further improve its performance. The framework can train an agent for a given problem with various parameters. As a use case, it is presented as a static obstacle avoidance maneuver. the proposed framework was tested on an automotive proving ground with the geometric constraints of the ISO-3888-2 test. The results proved its real-time capability and performance compared to human drivers’ abilities.