On Differential Drive Robot Learning Convex Policy with Application to Path-Tracking

This paper presents an experimental validation of a learning convex policy for path-tracking on a differential drive robot. An online implementation of the convex control policy (COCP) is provided in the ROS environment using the CVXGEN package that runs on the on-board computer in a real-time application. The control policies are trained in an off-board computer considering a stochastic kinematic description of the robot and using an approximate gradient method for a given cost-to-go metric function. The policy is validated through simulation and experimental evaluation. In addition, to certify the training efficacy, the experiment is also evaluated using the untuned policy. A discussion regarding trajectory errors is presented as well as final considerations for the solver and real-time concerns.