Neural Network Enabled Robotic Contact Measurement

Following the ideas put forth by industry 4.0, flexible manufacturing systems that make use of robots, sensors and artificial intelligence are gaining more and more relevance. While vision systems are fundamental for the implementation of autonomous robotic systems, they are often lacking in accuracy when the conditions are not ideal. A robotic contact measurement system could be a possible solution to measure with more accuracy portions on the surface of a work-piece where the vision systems may struggle to acquire accurately. Using a robot for contact measurement introduces uncertainties due to the dynamics of the robot and the geometry of the probe. In this work a data-driven adaptive filter is developed exploiting a trained inferential model in order to enable robots equipped with a force sensor to perform contact measurements. The model is trained using data acquired from a 3D printed part with small features representing small surface defects to be measured, as small as 0.2 mm. The result shows that the profile of the defect can be effectively reproduced through the use of model reducing the mean absolute error from 0.382 mm to 0.034 mm.