Edge Computing-based 3D Pose Estimation and Calibration for Robot Arms

Abstract

Industrial robots are widely used in current production lines, and complex pipeline processes, especially those with different assembly requirements, are designed for intelligent manufacturing in the era of industry 4.0. During the new crown epidemic, a large number of car companies used the production line to transform production of medical materials such as masks and protective clothing, which provided a strong guarantee for fighting the epidemic. In this scenario, a pipeline is often assembled from robotic arms from multiple suppliers. The traditional methods is complex and takes a lot of time. In this paper, we propose a novel deep learning based robot arm 3D pose estimation and calibration model with simple Kinect stereo cameras which can be deployed on light-weight edge computing systems. The light-weight deep CNN model can detection 5 predefined key points based on RGB-D data. In this way, when the assembly line composed of different robot arms needs to be reassembled, our model can quickly provide the robot’s pose information without additional tuning processes. Testing in Webots with Rokae xb4 robot arm model shows that our model can quickly estimate the key point of the robot arm.