Gaussian Process Model Predictive Admittance Control for Compliant Tool-Environment Interaction

Humans demonstrate remarkable dexterous manipulation ability by using tools to physically interact with the environments, e.g., refined polishing of workpieces using tools with soft foams. During the polishing process, modeling the tool-environment interaction is difficult because the robot may not have access to the physical properties of different compliant tools and environments (e.g., stiffness, shape, and friction). To address this challenge, a learning-based method has been employed to predict the contact model without explicitly knowing the intrinsic physical properties. Notably, the extrinsic contact is estimated using the intrinsic tactile sensing between the gripper and the tool in this work. Further, to ensure precise force control during the polishing process, a Gaussian process-based model predictive admittance control (GP-MPAC) scheme is developed with the estimated extrinsic contact. Finally, experiments have been conducted using a robot manipulator, and the results validate that the proposed method is effective while demonstrating its ability to handle the aforementioned complexities.