The TOR Agent:Optimizing Driver Take-Over with Reinforcement Learning

Various factors influence drivers’ response to Take-Over Requests in automated driving, and a wide range of designs have been proposed to improve transitions. Still, little research has investigated how systems could deliver Take-Over Requests at the best moment in time. In this paper, we sketch the idea of a reinforcement learning agent that learns to deliver Take-Over Requests at the right time so that drivers’ performance gets optimized, which could help to increase driving safety. We implemented such a system in Unity to evaluate this approach using a simple driver model. Our agent receives coordinates of the upcoming road segment and learns to deliver a Take-Over Request at an appropriate moment within a short time window. The reward function is composed to minimize the lateral deviation in the subsequent phase of manual driving. The initial results obtained are promising, and we will evaluate the concept with real human users soon.