High-Level Decision Making for Automated Highway Driving via Behavior Cloning

Abstract

Automated driving systems need to perform according to what human drivers expect in every situation. A different behavior can be wrongly interpreted by other human drivers and cause traffic problems, disturbances to other road participants, or in the worst case, an accident. In this paper, we propose a behavior cloning concept for learning high-level decisions from recorded trajectories of real traffic. We summarized and gave a clear definition of the main features that affect how humans make driving decisions. Some other approaches rely on complex neural networks where their decisions are impossible to understand. Due to the importance of the decision making module, we produce safe human-like behavior which is transparent to humans and easy to track. Furthermore, the learned policy is not overfitting to the limited training data and generalizes well to multi-lane scenarios with arbitrary speed limits and traffic density, which is strengthened by the successful application of merging policy on exiting. Simulation evaluations show that our learned policy is able to handle the intention uncertainty of surrounding agents, and provide human-like decisions in the sense of well-balanced behavior between efficiency, comfort, perceived safety, and politeness.