Speed-Up of Nonlinear Model Predictive Control for Robot Manipulators Using Task and Data Parallelism

Abstract

The repetitive evaluation of computationally expensive functions in the objective and constraints represents a bottleneck in the solution of the underlying optimal control problem (OCP) of nonlinear model predictive controllers (MPC) for robot manipulators. We address this problem by exploiting the parallel evaluation of such functions within the execution of a first-order and a second-order OCP solution algorithm, such as the proximal averaged Newton-type method for optimal control (PANOC) and the sequential convex quadratic programming (SCQP) method, respectively. The use of task parallelism with multicore executions and data parallelism with single-instruction-multiple-data (SIMD) instructions is shown to effectively reduce the solution time of the underlying OCP so that the satisfaction of real-time constraints in the deployment of MPC for robot manipulators can be achieved.