Speaker
Description
Optimal control problems in dynamic environments include the challenge that one has to make fast decisions without risking collisions. This can be illustrated with an example in which the goal is to drive on the racetrack without any collision with other road users. For this example, in [1], we introduced a hybrid structure, which consists of a Reinforcement Learning approach, which is used as a trajectory planner, and a solution approximation of a simplified optimal control problem in order to actually steer the dynamical system. This approach allows to make fast decisions for a fast changing environment. Unfortunately, the decision based on the Reinforcement Learning algorithm does not give any guarantees that all possible collisions are avoided, which is a typical problem of data based approaches. In contrast, in [2], a classical Reinforcement Learning approach is equipped with a funnel controller, which leads to new safety aspects. Nevertheless, in this approach a safe reference trajectory has to be available, which is often - like in the racetrack example - not the case.
In this talk, we will present a combination of these approaches. As an underlying example, we focus on the “driving on the racetrack”- example, where a safe controller in a fast changing environment is needed. We use Reinforcement Learning for trajectory planning, which is now safeguarded by a funnel controller nearby obstacles. Therefore, since it is not possible to compute a safe universal reference trajectory in advance, we define a safety trajectory for each obstacle that leads around the obstacle.
[1] S. Gottschalk, M. Gerdts and M. Piccinini: Reinforcement Learning and Optimal Control: A Hybrid Collision Avoidance Approach. In Proceedings of the 10th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2024), pages 76-87, 2024.
[2] S. Gottschalk, L. Lanza, K. Worthmann and K. Lux-Gottschalk: Reinforcement Learning for Docking Maneuvers with Prescribed Performance. 26th International Symposium on Mathematical Theory of Networks and Systems MTNS 2024, pages 196-201, 2024.
