Rudolf Reiter

SYSCOP

Wednesday, April 06, 2022, 10:45

online

We present an approach to plan trajectories in a race-car environment and propose a setting of a high-level reinforcement learning (RL) policy together with a low-level model-predictive control (MPC) formulation.
The RL policy adapts the cost function of a progression maximizing nonlinear program to achieve both, strategic and safe behavior. 
An interaction is considered with two simulated race-car agents, which are competing on a race track. We extend previous work with a novel RL/MPC interface and safety-guarantees. Our empirical simulation results show that the algorithm successfully trains safe and strategic high-level policies, providing a framework for interactive control.

Zoom:

https://uni-freiburg.zoom.us/j/62791737415
Meeting-ID: 627 9173 7415
Password: syscop2021