펙인홀 전략 학습을 위한 데이터 주도 물체 중심 학습 프레임워크

Abstract

In this study, a data-driven object-centric learning framework was developed for dealing with various shapes of pegs and holes in the peg-in-hole task. This framework enables efficient data sampling in simulation environments comprising pegs and holes of different shapes. A small regression model with minimal computational costs can be trained for inserting the peg into the hole. The model takes the contact force generated by the peg and the hole as input and infers the relative posture of the peg and the hole, moving the peg to minimize the relative pose. Peg-in-hole experiments were conducted using three shapes of pegs and holes, and the well-known spiral search strategy was compared with the proposed data-driven peg-in-hole strategy learning model. The proposed strategy outperformed the spiral search strategy in all three shapes of peg-in-hole tasks. Finally, the inference model was applied to a real-world environment, and a success rate of 73.3% was achieved in the one-degree-of-freedom peg-in-hole experiment.