Subsequent projects | Special projects

Start of funding 01.01.2021

Enhancing Safe Predictive Control with Learning-based Methods for Automated Vehicles

PD Dr.-Ing. habil. Dirk Wollherr
Technische Universität München
Chair of Automatic Control Engineering

Prof. Dr. Francesco Borrelli
University of California, Berkeley
Department of Mechanical Engineering


In recent years, model predictive control has proved to be a suitable control method to plan trajectories for automated vehicles. Uncertainty within the environment is accounted for by probabilistic constraints, referred to as stochastic model predictive control (SMPC). While probabilistic constraints allow efficient trajectory planning, a small probability of collision remains. Recently, however, a stochastic model predictive control algorithm has been developed that guarantees safety by planning fail-safe backup trajectories.

The goal of this research project is to extend this safe SMPC method. First, the efficiency of the developed safe model predictive controller is enhanced with learning-based model predictive control methods. This combination will provide advantages for currently challenging situations, such as icy and snowy conditions. Additionally, a test vehicle allows validating the simulations of the proposed method in experiments.

Final report:
In recent years, model predictive control has proved to be a suitable control method to plan trajectories for automated vehicles. As part of the project, two methods were further developed that enable safe planning for vehicles.

The first method allows a general backup of control approaches. For this purpose, a fail-safe backup controller is used, which becomes active whenever the desired control inputs would lead to future unsafe behavior. This allows the desired controller to be used as long as possible, but to have the safety guarantee of the fail-safe backup controller in safety-critical scenarios. This control concept is compatible with prediction-based and learning-based controllers and applicable to autonomous vehicles.

The second method deals with the case when assumptions made are invalid, for example, when another vehicle disregards traffic rules. In such cases, it is often not possible to guarantee safety. In the case of vehicles, the goal is then to minimize the probability of collision. For this purpose, we have developed a model predictive controller that minimizes the probability that constraints are violated.

There are now two possible next steps to be taken. On the one hand, a vehicle validation will provide insights where there is room for improvement for the two methods. On the other hand, a combination of the two developed methods for vehicles is of interest. This enables safe and efficient planning if all assumptions are met, and allows minimizing the collision probability if assumptions are violated.