The configuration of a two-wheeled vehicle, such as Segway, cannot be stabilized by continuous and time-invariant state feedback due to its non-holonomic constraints. Because of the nonlinear nature of the nonholonomic constraints, the realization of a model predictive control (MPC) for this class of vehicles is a difficult task.
This paper proposes a MPC method that can achieve long prediction horizon and quick computation. At the first step, the optimization of an input (i.e., velocity and steering) sequence is formulated as a graph search problem by restricting the inputs to discrete values. Next, in the second step, the optimized control result is learned by machine learning method, such as SVM.
A longer horizon MPC compared to that with nonlinear optimization can be realized. The advantages of the proposed method are demonstrated with simulation and experimental results.