This study proposes a multitask preferential Bayesian optimization (PBO) method for admittance learning in physical human-robot collaboration (HRC). The method enables automatic adjustment of admittance parameters by robots based on human preference feedback, thereby improving comfort and efficiency in cooperative tasks. Conventional optimization approaches in HRC often rely on performance metrics that inadequately capture subjective human preferences. To address this limitation, the proposed approach integrates pairwise preference learning within a Bayesian optimization framework, allowing the robot to learn optimal admittance parameters through intuitive human comparisons instead of numerical evaluations. Furthermore, this study extends conventional PBO into a multitask setting, where information from previously learned tasks or different human operators is shared through a multitask Gaussian process model with a covariance structure that incorporates task correlations. This enables the optimization process to start from prior knowledge, significantly reducing the number of required trials and improving sample efficiency. Experimental validation was conducted using a 7-DOF robot manipulator to perform point-to-point and path-following tasks with multiple operators. The results demonstrated that multitask PBO achieved faster convergence than conventional PBO, even when datasets from related tasks differed in task type or operator. Additionally, the findings indicate that learned damping parameters effectively reflect individual manipulation preferences and enhance cooperative task performance.
YASHIMA et al. (Thu,) studied this question.