To Support the Driver in Stabilizing the Powerslide

Machine Learning (ML) focuses on developing algorithms that enable computers to learn from data in order to make predictions or decisions. In this context, learning refers to the process of iteratively adjusting the parameters of a deep neural network to satisfy an optimality criterion best possibly. These networks are able to approximate complex, non-linear mappings between input and output data that may not be easily captured by traditional physics-based models.

Controlling a vehicle in extreme driving conditions, such as the powerslide, where large sideslip angles together with large traction forces and large negative steering angles of the handwheel occur, is challenging. To assist the driver, the aim of the investigated Advanced Driver Assistance System (ADAS) is to stabilize the unstable motion of the vehicle in powerslide condition, and the driver may focus on the path tracking task only. Thus, the presence of a human driver has to be considered by the ADAS – the controller has to account for the individual human driver control characteristics.

The controller of the data-driven approach is trained in a simulated environment where a nonlinear vehicle model and a human driver model are considered. While the driver model focuses on tracking the path, the controller’s objective is to stabilize the vehicle’s powerslide by applying torques to the front and rear axle. Like a cost function in optimal control, the reward function in reinforcement learning influences the control behaviour: The central objective is to find an optimal control policy that maximizes the sum of discounted rewards.

Project Goals

  • Training and evaluating a powerslide controller in a simulation environment.
  • Applying the controller on a real-world vehicle and minimizing the sim-to-reality gap.
  • Fine-tune the controller with real-world measurements for robust control on different and changing road surfaces and different types of human drivers.
Basic reinforcement learning interaction loop

Figure 1: Reinforcement Learning to support the driver at powerslide motion

Researchers

Project Partner

Ao.Univ.Prof. Dipl.-Ing. Dr.techn. Manfred Plöchl

University Lecturer, Research Unit of Technical Dynamics and Vehicle System Dynamics

Send email to Manfred Plöchl

Univ.Prof. Dipl.-Ing. Dr.techn. Johannes Edelmann

Head, Research Unit of Technical Dynamics and Vehicle System Dynamics

Send email to Johannes Edelmann