TY - GEN

T1 - Transfer learning for accurate modeling and control of soft actuators

AU - Wiese, Mats

AU - Runge-Borchert, Gundula

AU - Cao, Benjamin Hieu

AU - Raatz, Annika

N1 - Funding Information: 1All Authors are with the Institute of Assembly Technology, Leibniz Uni-versität Hannover, Germany wiese@match.uni-hannover.de Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 405032969 Fig. 1. Soft pneumatic actuator under pressure. Three integrated pressure chambers allow the soft pneumatic actuator to bend in three dimensional space.

PY - 2021

Y1 - 2021

N2 - The adaptability and inherent safety of soft material robotic systems offer great potential for applications in which their rigid counter parts reach their limits in terms of flexibility and safety. The soft materials used in these systems allow for a safe interaction between humans and robots. Despite advances in the development of soft robots in the recent years, for them to step into application, more research needs to be conducted in the field of accurate modeling and control. For model-based design, path planning, and control computationally efficient models need to be developed that are able to capture the often highly nonlinear deformation behavior of soft actuators. Our previous research showed that artificial neural networks (ANN) are a powerful tool for representig an actuator's nonlinear kinematics, while at the same time they are computationally efficient. In this article, we propose a transfer learning scheme for minimizing the effort of generating realworld data for neural network training.We showed that the generation of 50 real-world data pairs is sufficient to train an ANN that has a mean accuracy of less than 0.6% with respect to initial actuator length. The resulting ANN is applicable to open and closed loop kinematic control.

AB - The adaptability and inherent safety of soft material robotic systems offer great potential for applications in which their rigid counter parts reach their limits in terms of flexibility and safety. The soft materials used in these systems allow for a safe interaction between humans and robots. Despite advances in the development of soft robots in the recent years, for them to step into application, more research needs to be conducted in the field of accurate modeling and control. For model-based design, path planning, and control computationally efficient models need to be developed that are able to capture the often highly nonlinear deformation behavior of soft actuators. Our previous research showed that artificial neural networks (ANN) are a powerful tool for representig an actuator's nonlinear kinematics, while at the same time they are computationally efficient. In this article, we propose a transfer learning scheme for minimizing the effort of generating realworld data for neural network training.We showed that the generation of 50 real-world data pairs is sufficient to train an ANN that has a mean accuracy of less than 0.6% with respect to initial actuator length. The resulting ANN is applicable to open and closed loop kinematic control.

UR - http://www.scopus.com/inward/record.url?scp=85108702104&partnerID=8YFLogxK

U2 - 10.1109/RoboSoft51838.2021.9479300

DO - 10.1109/RoboSoft51838.2021.9479300

M3 - Conference contribution

AN - SCOPUS:85108702104

SN - 978-1-7281-7714-4

SP - 51

EP - 57

BT - 2021 IEEE 4th International Conference on Soft Robotics, RoboSoft 2021

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 4th IEEE International Conference on Soft Robotics, RoboSoft 2021

Y2 - 12 April 2021 through 16 April 2021

ER -