Calibration by diffeomorphisms of robot manipulator kinematics: a novel approach

Abstract

Our paper presents a nonparametric data-driven technique that can enhance the accuracy of robot kinematics models by reducing geometric and nongeometric inaccuracies. We propose this approach based on the theory of singular maps and the Large Dense Diffeomorphic Metric Mapping (LDDMM) framework, which has been developed in the field of Computational Anatomy. This framework can be thought of as a method for identifying nonlinear static models that encode a priori knowledge as a nominal model that we deform using diffeomorphisms. To tackle the kinematic calibration problem, we implement Calibration by Diffeomorphisms and obtain a solution using an image registration formalism. We evaluate our approach via simulations on double pendulum robot models, which account for both geometric and nongeometric discrepancies. The simulations demonstrate an improvement in the precision of the kinematics results for both types of inaccuracies. Additionally, we discuss the potential application of physical experiments. Our approach provides a fresh perspective on robot kinematics calibration using Calibration by Diffeomorphisms, and it has the potential to address inaccuracies caused by unknown or difficult-to-model phenomena.