Nonlinear model predictive control hardware implementation with custom-precision floating point operations

Abstract

Model predictive control (MPC) based techniques have found many applications both in academia and in industry. Its reach, however, may not be compared to classical control techniques due to e.g. the difficulty of solving an optimization problem at each sampling interval with real-time requirements. Most of the efforts to make the application of MPC viable address this problem with more efficient solvers. This paper, in contrast, proposes a new approach for a real-time MPC solution by mapping an approximate off-line solution into an artificial neural network in a FPGA (Field Programmable Gate Array). We implemented a radial basis function artificial neural network on a low cost FPGA using custom precision floating point operations and tested the control on a single-link robotic manipulator. The amount of time used to calculate the control action at each time instant is in around one microsecond. The comparison between the offline and the approximate solution shows the soundness of the idea. We provide an analysis of hardware usage and execution time in order to achieve the best compromise considering the precision for a given application.