We will follow a data-driven approach to achieve human-like driving styles with human-level adaptability and personalization to the human driver/passenger. We will estimate driving controllers from collected experience and we will extract a library of different maneuvers from demonstrated data. We will use the maneuver library to plan the trajectory of the car by switching between different maneuvers. We will also use optimal control and reinforcement learning techniques to improve the single maneuvers such that the maneuvers generalize to unseen situations and possibly even outperform the human drivers. In particular, we will concentrate on learning to resolve dangerous situations such as avoiding an unexpected obstacle. An important research question for using maneuver libraries is how to switch between maneuvers. The system should produce as little number of switches as necessary to generate a smooth driving behavior. Moreover, we need to incorporate high-dimensional sensory input from the environment in the switching decision. To do so, we will investigate the use of deep learning techniques.
Contact: Gerhard Neumann (PI)