A New Approach to Enhance Artificial Intelligence for Robot Picking System using Realistic Digital Twins Tool

Abstract

Robotic bin picking system (RBP) with Artificial Intelligence (AI) has been widely used in different applications for learning the features of new workpieces and detecting their respective coordinates through a camera.  One of the biggest bottlenecks of AI is the need for a vast amount of labeled data for sufficient training of the AI model. If either the quantity of the data is insufficient or the quality of the labeling is unstable, problems will arise when training and testing the AI. Also, many of the robotic picking systems will use a vacuum as the main gripper because the soft vacuum cup can easily adapt to the workpiece, however because of the different speeds of the handling, weights of the workpieces, and the air pressure exerted by the vacuum cup there is diversity in the behavior of the grasping. Thus, it is challenging to find a digital twin system to verify and analyze RBP to improve its real-world performance.

To resolve the first of the problems, we propose an early deployment method for automatically generating diverse data with domain randomization and an auto-picking point annotation system for labeling the data. For the second problem, we employ the usage of mathematical formulas to calculate and approximate the gripper reactions using various parameters.

Our proposed system will train and implement an AI for RBP using its own generated dataset in an early stage. Furthermore, we will test its performance in the simulator and real-world with a vacuum gripper to validate our system and formulas.