Adapting to unknown environment conditions in autonomous racing competition using Iterative Learning Control

Scaled-down autonomous racing presents a challenging and highly competitive setting, where achieving minimal lap times requires overcoming a variety of challenges, including the selection of an optimal racing line. In this paper, we address the optimal racing line problem together with environment awareness. We propose an improved racing line encoding technique based on an approximate homeomorphic transformation that maps track segments to unit squares used by an Evolutionary Algorithm. The quality of a candidate racing line is evaluated using a drive simulation based on a single-track vehicle model. Using Covariance Matrix Adaptation Evolution Strategy and isochronal segmentation, our method outperforms state-of-the-art approaches. Adaptation to unforeseen environment conditions is achieved through an online Iterative Learning Control paradigm that learns the discrepancy between the planned and realized racing line and accordingly adjusts the control reference of the Pure Pursuit algorithm used for racing line tracking. Our experimental evaluation demonstrates that our algorithm is able to adapt to unknown and spatially varying environment parameters, such as friction, to improve racing performance. We show that using the environment awareness can reduce lap time by an additional 14%.