Variable series elastic link: Advancing stiffness controllability in robot manipulators

Abstract

In an ageing population the need for assistive robotics has a great potential to address issues around the increasing demand for nursing and caregiving. Areas that robots may play a role are in helping with the activities of daily living (ADL) and dressing is the focus of this paper. Successful integration of these robots into society will require careful consideration of factors such as safety and interaction. We believe that these systems should be able to predict the user’s intention for maximum safety and task efficiency. Using data collected from human-human interaction (HHI) experiments, features were prepared and assessed for importance and models were trained to classify the dressing task segment and which end effector to move; left, right or both simultaneously. Long short term-memory networks (LSTM) were explored to predict these outcomes one timestep ahead. The networks were assessed against a variety of hyper-parameters including the depth of the hidden layers. The models show promise for correctly classifying task segment based on user pose, with the best test accuracy >95%.