Hybrid Mechanical and Data-driven Modeling Improves Inverse Kinematic Control of a Soft Robot

Abstract

Feed-forward control relies on accurate knowledge about the controlled plant, e.g. models of manipulator kinematics or dynamics. However, for many plants, mechanical models do not capture all aspects of a plant or the plant's intrinsic properties, e.g. soft materials, do hardly allow for exact and efficient mechanical modeling. In this context, machine learning is a suitable technique to extract non-linear plant models from data. The paper shows that feed-forward control based on inversion of a hybrid forward model comprising a mechanical model and a learned error model can significantly improve accuracy. The proposed approach is demonstrated for inverse kinematic control of a redundant soft robot with a hybrid model that is constructed from continuum kinematics together with an efficient neural network error model.