3D Segmentation of Bin Picking by Domain Randomization

Bin picking is a fundamental task in robotic manipulation and widely used in industrial manufacturing. Despite the significant progress in applying 3D visual guidance to bin picking, accurate segmentation of one object from a large number of parts in the presence of occlusion and stacking remains challenging. To tackle this problem, recent research has proposed general deep learning-based approaches. However, manual annotation requires significant effort, which we aim to address in this paper by implementing automatic annotation via domain randomization. Domain randomization is one of Sim2Real’s approaches to validate real data through annotation in a simulated environment.In this paper, we propose a bin-picking segmentation method by combining 2D object detection and a deep learning based 3D semantic segmentation model. We also propose a novel domain randomization method for both object detection and semantic segmentation. In addition, we train the bin-picking deep learning model with proposed method and validate it on test data obtained from real RGB-D camera. The experimental results presented in this paper demonstrate the advanced capabilities of our hybrid method. We have achieved high-precision segmentation,even in challenging scenarios characterized by substantial stacking, disorder, and the presence of noise, which validate the effectiveness of our approach.