Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 06, 2021 - June 08, 2021
Peg-in-hole task has been studied as a benchmark task for robotic assembly. It involves two main phases: search phase and insertion phase. This paper proposes a method that uses Reinforcement Learning(RL) to achive search phase in the peg-in-hole task. In this method, the stiffness matrix for admittance control is generated online. The method uses a visual sensor to determine the relative position of peg and hole, and selects an appropriate stiffness matrix model. By using visual sensor, this method has two advantages: it reduces the number of learning episodes and speeds up the search process. The two advantages of the proposed method were verified by peg-in-hole task using a 6-DOF manipulator.
