Towards Predictive Path-Following Control using Deep Neural Networks and Path Primitives

Model Predictive Path Following is an attractive solution for motion control of mobile robots and other autonomous systems. It requires solving a non-convex constrained optimization problem at each sampling period. We focus on a general approach to approximate Model predictive Path-Following Control (MPFC) for a mobile robot using a Deep neural network (DNN) that learns a set of base MPFC representations via path primitives. We show that using simple algebraic operations, any path can be followed with reasonable accuracy, and we illustrate how to apply this approach for paths described by linear segments and parabolic blends that can be generated by a robotic path planning algorithm. Compared to the computational requirements of MPFC, our proposed approach requires significantly less memory and the execution speed is two orders of magnitude faster. This makes our approach suitable for microcontroller implementation, with only a small degradation of the path-following accuracy compared to online MPFC.