Learning the Elasticity of a Series-Elastic Actuator for Accurate Torque Control
Bingbin Yu, José de Gea Fernández, Yohannes Kassahun, Vinzenz Bargsten
Editors: Salem Benferhat, Karim Tabia, Moonis Ali
In Advances in Artificial Intelligence: From Theory to Practice: 30th International Conference on Industrial Engineering and Other Applications of Applied Intelligent Systems, IEA/AIE 2017, Arras, France, June 27-30, 2017, Proceedings, Part I, 27.6.-30.6.2017, Arras, Springer International Publishing, pages 543-552, Université Artois, Jun/2017. ISBN: 978-3-319-60042-0.
Series elastic actuators (SEAs) have been frequently used in torque control mode by using the elastic element as torque measuring device. In order to precisely control the torque, an ideal torque source is critical for higher level control strategies. The elastic elements are traditionally metal springs which are normally considered as linear elements in the control scheme. However, many elastic elements are not perfectly linear, especially for an elastic element built out of multiple springs or using special materials  and thus their nonlinearities are very noticeable. This paper presents two data-driven methods for learning the spring model of a series-elastic actuator: (1) a Dynamic Gaussian Mixture Model (DGMM)  is used to capture the relationship between actuator torque, velocity, spring deflection and its history. Once the DGMM is trained, the spring detection can be estimated by using the conditional probability function which later is used for torque control. For comparison, (2) a deep-learning approach is also evaluated which uses the same variables as training data for learning the spring model. Results show that the data-driven methods improve the accuracy of the torque control as compared to traditional linear models.
series-elastic actuators, nonlinear springs, DGMM, deep learning, torque control