An Industrial Bin Picking Framework for Assembly Tasks

Abstract

The majority of current bin picking systems, designed for industrial parts, cannot be directly oriented to the downstream task after grasping. This research presents a grasping framework that addresses this challenge by incorporating pose estimation of parts in cluttered bin environments and the targeted design of robot end-effector grippers. This approach ensures that the pose of the part on the gripper is known and fixed, enabling successful assembly tasks in various scenarios. To train an object pose estimation network, we propose a system for generating a dataset of industrial parts using model rendering within a physics engine. We analyze the geometric features of the parts, and further design a gripper, to achieve the grasping strategy. Results demonstrate that for a single known industrial part, the minimum grasping success rate is 91.4% in simulated robot experiments, and the assembly success rates in different scenarios based on this framework exceed 80%. Our framework offers valuable guidance for the deployment of robotic grasping.