Feedforward Control Approaches to Bidirectional Antagonistic Actuators Based on Learning

Abstract

Safe physical human-robot interaction is a decisive feature in wider adaptation of robots in homes and factories. To that end, a lot of researchers consider new actuation mechanisms and particularly Variable Stiffness Actuators (VSAs) which contribute to robot safety, but also to increase energy efficiency and outperforming rigid actuators in repetitive tasks. However, advantages of VSAs come with their price – issues in design and control of such multivariable non-linear systems. Novel approaches and methods in soft computing methods such as machine learning and neural networks are opening new horizons in VSA control. In this paper, a comparative analysis is carried out between the neural network feedforward control and locally weighted projection regression as a technique for model learning of bidirectional antagonistic VSA – qb move maker pro. Set of measurement is used to create mapping between two motor positions as inputs and measured actuator position and estimated stiffness as outputs. Comparative analysis of the two different approaches for feedforward control observing performances in open loop control, followed by closed loop testing with a simple feedback regulator for fine tuning. Learning techniques result in robust and generalized models that can predict required inputs in ordered to achieve good output tracking.