Automatic grasp generation and improvement for industrial bin-picking

Abstract

This paper presents work on automatic grasp generation and grasp learning for reducing the manual setup time and increase grasp success rates within bin-picking applications. We propose an approach that is able to generate good grasps automatically using a dynamic grasp simulator, a newly developed robust grasp quality measure and post-processing methods. In addition we present an offline learning approach that is able to adjust grasp priorities based on prior performance. We show, on two real world platforms, that one can replace manual grasp selection by our automatic grasp selection process and achieve comparable results and that our learning approach can improve system performance significantly. Automatic bin-picking is an important industrial process that can lead to significant savings and potentially keep production in countries with high labour cost rather than outsourcing it. The presented work allows to minimize cycle time as well as setup cost, which are essential factors in automatic bin-picking. It therefore leads to a wider applicability of bin-picking in industry.