Driving at an unsignalized intersection is a complicated driving scenario that requires both safety and traffic efficiency. In the intersection, the driving policy does not simply maintain the safe distance to all vehicles, but should pay more attention to the vehicles that are likely to cross with the ego vehicle and make its decision considering their intentions. Our goal is to train an attention-based driving policy for handling intersection scenarios using deep reinforcement learning. By leveraging the attention, our policy is able to learn how to focus on spatially and temporally more important features in its egocentric observation and perform complex driving strategies at the congested intersection environment.
We transfer the policy model trained in a high fidelity simulator to a full-scale vehicle system, and conduct experiments to evaluate our model in simulated and real-world environments. Our model successfully performs various intersection scenarios even with noisy sensory data and delayed response.