6DoF Pose-Estimation Pipeline for Texture-less Industrial Components in Bin Picking Applications

Abstract

Over the next few years, autonomous robots and functionalities are expected to gain increased importance for the shop floor. Perception and the derivation of autonomous behavior is of crucial importance in this context. We present a combined object recognition and pose estimation pipeline to generate pose estimates with six degrees of freedom (6DoF) for bin picking, specifically targeting the suitability for challenging scenarios with texture-less, metallic parts in industrial environments. The pipeline is based on open source algorithms, combining Convolutional Neural Networks (CNNs) and feature-matching methods to create an effective 6DoF pose estimate. We evaluate our approach on several industrial components using a articulated arm robot to guarantee a high level of comparability during the different measurement runs. We further quantify the results using known error metrics for pose estimation, compare the results to established approaches and provide statistical insight into the achieved outcomes to assess the robustness and reliability.