An adaptive inexact-Jacobian SQP method for predictive path-following

Niels van Duijkeren

KU Leuven, Belgien

Tuesday, June 14, 2016, 11:00

Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany

Path-following control refers to the task of following a geometric reference without a timing law assigned to it. Hence, unlike in reference-tracking control, the geometric path does not explicitly state when to be where on the reference curve and leaves this degree of freedom as a decision variable for the controller. Predictive approaches to path-following, which can be written as NMPC problems, have been proposed for their ability to consider input and state constraints. A reformulation of the path-following NMPC problem, to obtain a spatial horizon instead of a temporal horizon, has been proposed before in application to path-following control for a robot arm. Solution times of the RTI-scheme implementation using the ACADO Code Generation Toolkit has shown to be too slow for a real-time implementation for considering all degrees of freedom. This talk discusses one approach to mitigate the computational performance issues through the application of SQP with inexact Jacobians. An adaptive scheme is proposed to initialize sensitivities as they appear in the prediction horizon and to only compute a limited number of sensitivities each iteration according to an update criterion. The algorithm is validated in simulation experiments for an application to a robot arm.