Deep reinforcement learning-based antilock braking algorithm

Abstract

Tires play an important role in the performance of vehicular safety systems. Antilock braking system is one of the most important active safety systems that interacts with the tires. Unlike a variety of existing algorithms which are tuned to a specific tire, this research proposes a model-free reinforcement learning-based control algorithm which can adapt to changing tire characteristics and there by effectively utilising the available grip at tire–road interface. The simulation model, consisting of brake actuator dynamics, transportation delays, tire relaxation behaviour, vehicular longitudinal and pitch dynamics, is trained using more than 350,000 random tires. To reduce training time, a parallelisation architecture has been proposed which distributes learning and simulation tasks to different CPU cores. Finally, a conditional variance-based sensitivity analysis with over twelve thousand tires indicate improved grip utilisation at tire–road interface and decreased sensitivity of stopping distance on tire nonlinearity compared to literature version of Bosch algorithm.