Teaching a vehicle to autonomously drift: A data-based approach using Neural Networks

This paper presents a novel approach to teach a vehicle how to drift, in a similar manner that professional drivers do. Specifically, a hybrid structure formed by a Model Predictive Controller and feedforward Neural Networks is employed for this purpose. The novelty of this work lies in a) the adoption of a data-based approach to achieve autonomous drifting along a wide range of road radii and body slip angles, and b) in the implementation of a road terrain classifier to adjust the system actuation depending on the current friction characteristics. The presented drift control system is implemented in a multi-actuated ground vehicle equipped with active front steering and in-wheel electric motors and trained to drift by a real test driver using a driver-in-the-loop setup. Its performance is verified in the simulation environment IPG-CarMaker through different open loop and path following drifting manoeuvres.