Semi-parametric identification of manipulator dynamics in a time-varying environment

Abstract

A recent trend in robotics is aimed at the cooperation between human and robot. This has led to an increased development of collaborative robot manipulators. Typical characteristics of collaborative robots are their user-friendly and lightweight design, innovative compliant mechanics, the implementation of various safety features and advanced control capabilities. These characteristics enable humans to work alongside the manipulator or interact with it. The implementation of passive compliant components such as springs and pneumatics have a beneficial effect on the level of safety for the operator. However, the added complexity often has a negative influence on the degree to which an accurate description of the system dynamics can be derived. Furthermore, the lightweight design and increasing payload-to-weight ratio amplify the effect of exogenous alterations to the system, such as attaching an object to the end effector. The work in this thesis is aimed at obtaining an accurate description of the system dynamics for control purposes. In doing so, special attention is given to dealing with instantaneous time-varying phenomena. An online semi-parametric approach is used to produce a valid description of the inverse dynamics of the considered system. The method consists of a non-parametric part which is described using Gaussian process regression (GPR) and a parametric part for which the parameters are identified using an extended Kalman filter (EKF). In this thesis, instantaneous system changes are introduced by attaching an unknown object to the end effector of the manipulator. The EKF implementation is specifically aimed at rapidly compensating for the response induced by this object. The GPR is used to compensate for remaining modeling errors. The performance of the proposed methods is evaluated in simulation. The semi-parametric description achieves high modeling accuracy, fast adaptation to instantaneous system changes and reasonable generalization capabilities. Implementing the proposed solution in real-time applications requires additional research on the subject of online GPR.