Object-Level Stiffness Control for the Dexterous Hand in Limited Perception With Grasping Learning and Data Fusion

Object-level stiffness control is a method aiming to derive specific stiffness for an object while grasping it with dexterous hands. However, this usually needs to be supported by sufficient sensing information, such as joint torque or 3-D force of fingertips. Therefore, the accurate stiffness control is a challenge for dexterous hands that lack these sensors. Achieving this goal with limited perception relies on estimating the interaction force between the grasped object and the environment. In order to achieve the better estimation of interaction force and to achieve better object-level stiffness control, this article proposes a new framework that combines grasping feature learning and data fusion. In this framework, a grasping feature learning method based on support vector regression is proposed to model the mapping relationship between the grasping features and the fingertip force sensor confidence. Furthermore, the Kalman filter algorithm is used to complete the data fusion of the force sensing information and the confidence information output from the grasping learning model, so as to obtain the optimal estimation of the interactive force. Finally, by designing fingertip controllers with switchable states on the three-fingered dexterous hand, a precise object-level stiffness control process is realized.