Robotic Learning of Haptic Skills From Expert Demonstration for Contact-Rich Manufacturing Tasks

Abstract

We propose a learning from demonstration (LfD) approach that utilizes an interaction (admittance) controller and two force sensors for the robot to learn the force applied by an expert from demonstrations in contact-rich tasks such as robotic polishing. Our goal is to equip the robot with the haptic expertise of an expert by using a machine learning (ML) approach while providing the flexibility for the user to intervene in the task at any point when necessary by using an interaction controller. The utilization of two force sensors, a pivotal concept in this study, allows us to gather environmental data crucial for effectively training our system to accommodate workpieces with diverse material and surface properties and maintain the contact of polisher with their surfaces. In the demonstration phase of our approach where an expert guiding the robot to perform a polishing task, we record the force applied by the human (Fh) and the interaction force (Fint) via two separate force sensors for the polishing trajectory followed by the expert to extract information about the environment (Fenv = Fh - Fint). An admittance controller, which takes the interaction force as the input is used to output a reference velocity to be tracked by the internal motion controller (PID) of the robot to regulate the interactions between the polisher and the surface of a workpiece. A multilayer perceptron (MLP) model was trained to learn the human force profile based on the inputs of Cartesian position and velocity of the polisher, environmental force (Fenv), and friction coefficient between the polisher and the surface to the model. During the deployment phase, in which the robot executes the task autonomously, the human force estimated by our system ( <^>Fh) is utilized to balance the reaction forces coming from the environment and calculate the force ( <^>Fh - Fenv) needs to be inputted to the admittance controller to generate a reference velocity trajectory for the robot to follow. We designed three use-case scenarios to demonstrate the benefits of the proposed system. The presented use-cases highlight the capability of the proposed pHRI system to learn from human expertise and adjust its force based on material and surface variations during automated operations, while still accommodating manual interventions as needed.