Learning Peg-in-Hole Assembly Skills from Demonstration Based on Geometric Constraints

Abstract

Existing Learning from Demonstration (LfD) methods that can learn the peg-in-hole assembly skills mainly adopts the impedance control strategy, and generally do not pay attention to the assembly skill segmentation problem. This scheme is susceptible to environmental noise and is difficult to learn stable and effective assembly skills. In this paper, a geometric constraints-based LfD method is proposed and applied to the peg-in-hole assembly task. By identifying geometric constraints from the motion trajectory and considering the contact force information, the demonstration dataset is first segmented into a series of different sub-skills. Hybrid Force/Position Control (HFPC) schema is adopted to model those skills. The selection matrix is identified according to the geometric constraints. The Gaussian Mixture Model (GMM)/Gaussian Mixture Regression(GMR) and quasiperiodic motion models are used to representing the reference trajectories. A switching mechanism is also learned to coordinate those HFPCs with different configurations to accommodate different sub-skills. The experimental results show that the proposed method can accomplish the learning and reproduction of assembly tasks stably and effectively.