The longitudinal controller of Autonomous Vehicles (AV) plays a key role in maintaining human comfort, road safety and compliance with road regulations. This study presents a novel data-driven framework for the customisation of longitudinal controllers based on human driving data. This framework facilitates the learning of driving styles of individuals and aims to improve longitudinal ride experience using minimal training data. Two architectures are proposed to realise the framework, both using supervised learning with expert human demonstration allowing the Feed Forward Neural Network (FFNN) to mimic human behaviour. In the first architecture, the output of the FFNN is limited to a safe and legal speed. The second approach uses a virtual vehicle-based solution to maintain safe and legal speeds. The performance of the proposed controllers is evaluated across various test scenarios. The results show that the proposed controllers achieve human-like driving behaviour and show adherence to the speed limits. The performance of the controllers is also evaluated in an Out-Of-Distribution (OOD) scenario where insufficient training data is supplied. The results show the capability of the controllers despite this challenge. The overall outcomes of the study show improvement in the human like motion and personalisation from AV longitudinal controllers and enhancement of the ride experience and enjoyment.
Gregory et al. (Sun,) studied this question.