An Intermittent Learning Algorithm for High-Speed Autonomous Driving in Unknown Environments

Abstract

An intermittent, model-free optimal control algorithm that enables an autonomous vehicle to track a non-predetermined trajectory at high speed is presented. The approach is bandwidth and energy efficient in that communication between actuators is limited to instances when it is needed rather than performing unnecessary periodic updates. We formulate the problem by properly augmenting the system and reference (trajectory) data, and then designing a triggering mechanism for the controller to work with a sampled version of the augmented states at some triggering instants. In order to obtain a model-free solution, we leverage a Q-learning framework with a zero-order hold actor network and a critic network to approximate the optimal intermittent controller and the optimal cost, respectively, resulting in appropriate tuning laws. Finally, we provide a numerical example of an ground vehicle driving autonomously at high-speed on a race track.