Learning end-to-end diagnostic policy for passive rehabilitation robot with deep learning
Konferenz: BIBE 2018 - International Conference on Biological Information and Biomedical Engineering
06.06.2018 - 08.06.2018 in Shanghai, China
Tagungsband: BIBE 2018
Seiten: 6Sprache: EnglischTyp: PDFPersönliche VDE-Mitglieder erhalten auf diesen Artikel 10% Rabatt
Ni, Peiyuan; Zhang, Wenguang; Hu, Kaixiang; Wang, Peifei; Cao, Qixin (State Key Lab of Mechanical Systems and Vibration, Shanghai Jiao Tong University, Shanghai, China)
For passive rehabilitation robots with low costs and simple systems, their control policies require to be adjusted manual-ly. Moreover, the state variables for the policies also require to be defined by human. Facing different patients with different symptoms, it is hard to provide the individualized treatments because it is a tedious work to adjust the parameters for the robots. With the development of deep learning, many research works focus on how to train the robot to learn different skills. We hope our passive robot has the ability to autonomously diagnose the patient and adjust the parameters by itself. Here we propose an end-to-end method with deep learning to help the robot to learn the experience from the therapists. We first train a supervised network to deal with all the expert’s data. Then we propose a policy fine-tuning algorithm to reduce the compound errors. Finally, our experiment shows that our robot can autonomously diagnose and adjust the parameters effectively.