Towards One Shot & Pick All: 3d-Oas, an End-to-End Framework for Vision Guided Top-Down Parcel Bin-Picking Using 3d Overlapping-Aware Instance Segmentation and Gnn

Abstract

Robotic Grasping and sorting is acknowledged as the most fundamental and significant manipulation task in industry. An ultimate goal of a Vision guided Robotic grasping system is to precisely and efficiently pick maximum objects with minimum shots and processing time of vision system. So far, applicable end-to-end perception and autonomously picking of hierarchically stacked objects has not been fully discussed in previous works. Especially, in the scenario of top-down parcel bin-picking where robots are required to perceive and pick up parcels from random stack. In this work, we focus on this challenging task by putting forward a novel end-to-end parcel bin-picking model termed 3D-OAS. Our proposal combines a 3D overlapping-aware instance segmentation and directed graph to describe the hierarchical structure of stacked objects from a top-down angle and a graph-neural-network is introduced to solve the optimal sorting orders. The experiment was conducted via a set of Vision guided Delta-Parallel robotic grasping system with a topdown RGB-D camera. Experimental Results proved the feasibility of our proposal, it could hierarchically segment stacked objects and solve sorting sequence with minimum one shot.

Keywords: Parcel bin-picking, vision guided robotic grasping, 3D overlapping-aware instance segmentation