Learning autonomous drift parking from one demonstration

This paper concerns with motion control of an all-wheel-drive (AWD) radio-controlled (RC) car in a setting of drift parking, which is made difficult by the stochastic, coupled dynamics. Motivated by human who can often learn complicated motion from few samples without any mathematical modeling, we propose an online learning algorithm for this task. The idea is to learn a coarse dynamics model from a single demonstration using regression, then to refine a parametric control policy by an online optimization via this model. At the core of the approach lies a local baseline method for the variance suppression on the gradient update to facilitate a fast learning. The algorithm makes use of our intuitive understanding in car driving to efficiently extract a coarse model from samples, and further leverages this coarse model to guide the update of control policy towards smaller state deviations during the trajectory following. Furthermore, the conditions on the performance under which this algorithm is guaranteed are derived and discussed. Finally, the superior performance of this algorithm is demonstrated in experiments.