Genetically generated bionic driver models for autonomous road vehicles

Abstract

For autonomous road vehicles, control is often divided into longitudinal and lateral control. This thesis focuses on lateral control driver models derived from a cognitive perspective. A genetic algorithm is used to generate driver models expressed in a domain-specific language. The project focuses on isolating perceptual cues. The objective function for the genetic algorithm is computed as the difference between the estimated steering angles and the observed steering angles in the vehicle. The recordings were captured from a Volvo XC90 driving a single scenario with an S-shaped test track at different speeds, and with different drivers. The resulting driver models are within 1-2 degrees of the recorded steering angles, and more significantly, the DSL sentences are very similar regardless of driver or speed, and stable between different runs. The project’s results show that the implementation of the genetically generated driver models is possible for lateral control. This genetic algorithm serves as a platform for the future inclusion of external factors affecting the dynamics of the vehicle. The identified model and parameters can be tested for representing a real-world driving case. Keywords: