Data-efficient human training of a care motion controller for human transfer assistant robots using Bayesian optimization

Abstract

Transferring is one of care motions for transferring a patient between a bed and a wheelchair, which is a heavy burden task for caregivers. Various kinds of transferring assistant robots have been proposed previously. However, designing comfortable transfer motions for individuals has not been intensively studied maybe due to its difficulty in modeling of such physical human-robot interactions and user's preference. In this paper, we propose a novel framework for generating comfortable transfer motion for individuals. We find the comfortable motion from the user feedback (scalar value) through human-robot physical interactions. To make such an approach practical, it is critical to reduce the number of required experiments with physical interactions between the user and the robot for reducing user's burden. We formulate the task as a black-box optimization problem of a unknown noisy function; Bayesian optimization, a data-efficient black-box optimization method, is utilized for quickly searching the optimal care motion. In our experiment, a dual-arm human transfer assistant robot is used for the raising upper body task in lift-up state. Experimental results demonstrate that our method can optimize a user-comfortable care motion controller for each individual efficiently. Moreover, the relationship between the user feedback and the sensor data associated with the robot care motions are analyzed using linear regression analysis.