A vision-based error compensation method for accurate path tracking in robotic trimming
Loading...
Date
2023-08-16
Authors
Tayaranian Marvian, Keyvan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Trimming is a common step in the fiberglass manufacturing process. During the trimming operation, outer edges and inner cutouts of molded fiberglass parts are removed to get the final part. This process is commonly performed manually by workers in many fiberglass manufacturing plants. Robotic automation of trimming processes is a challenging task due to the highly variable nature of fiberglass manufacturing. Fiberglass parts typically suffer from manufacturing inconsistencies and deformations, and rendering pre-defined offline robot programs is impractical.
To enable robotic automation of trimming processes, it is necessary for robots to detect and adjust to part variation. This research develops a methodology in which a fusion of vision and laser sensors along with advanced image processing and robot control techniques are used to automatically detect and accurately follow trimming paths on fiberglass parts.
A multi-stage real-time and offline error compensation framework is proposed. An external 3D camera and point cloud processing techniques are used to automatically detect trimming paths and generate target points to guide a robot. To improve the accuracy of the robot path, a 2D camera mounted on the robot is used to directly measure and correct the path deviation. A laser displacement sensor is also used to implement real-time height control, ensuring a constant distance between the onboard camera and the surface. A laser cross sensor is also utilized to measure and correct the orientation errors.
Moreover, a deep learning model is developed to improve the robustness of the path detection step. In comparative experiments, variants of the U-Net architecture and backbones with different hyperparameters are compared to find the best-performing model. A U-Net model with an Xception backbone is trained to be a classifier with an AUC value of 0.99 and 96.38% accuracy on test data. The developed model is tested on a sample fiberglass part using an industrial robot. The results show that errors can be reduced to less than 0.5 millimeters and 3 degrees, which meet the required tolerance in typical fiberglass manufacturing applications.
Description
Keywords
Robotic Trimming, Vision-Based Error Compensation, Path Tracking, Path Detection, Deep Learning