Damping profile learning for human-robot collaboration using Bayesian optimization with a task success rate model

In the realm of human-robot collaboration, impedance and admittance control are widely utilized techniques to regulate the interaction dynamics between humans and robots. This study proposes a novel approach to enhance the performance of human-robot collaboration by adapting the damping profile used in admittance control. The proposed method employs Bayesian optimization to appropriately adjust the damping profile, even when the cost function and unknown constraints are computationally expensive to evaluate, which is a common scenario in human-robot collaboration. Furthermore, the proposed approach can accommodate diverse types of constraints, enabling the incorporation of prior knowledge as constraints and accelerating the learning process. Specifically, constrained Bayesian optimization, which utilizes both Gaussian process regression and classification models, was implemented to learn damping profiles while considering task success rates. Additionally, a dynamic time warping technique was employed to handle task failure evaluations during the learning process, effectively mitigating the influence of outlying observations. Furthermore, prior knowledge regarding the damping profile is incorporated as a constraint to improve the learning performance. Extensive simulations and real-world experimental evaluations with a 7-DOF robot arm demonstrate that the proposed method can generate appropriate impedance parameters, exhibiting a substantial advantage over conventional impedance-learning methods in terms of learning performance.