A Multi-Camera Framework for Visual Servoing of a Collaborative Robot in Industrial Environments

A great part of today’s industries tends to invest on automatic machines that can replace or collaborate with humans in typical repetitive tasks. Despite their high motion and positioning precision, most of these industrial robots operate blindly, causing the working system to be poorly robust to even slight changes of the working conditions. A solution to such an issue might be to make the robots capable of readjusting their actions according to a perceptual feedback, in particular made of visual data. In this work we propose a multi-camera framework for the visual servoing of a collaborative robot that has to manipulate untextured industrial pieces. The robot is supposed to recognize the object of interest and reach it with its end-effector. We adopt a multi-camera approach that overcomes typical issues related to single-camera schemes. The system contains an object recognition module that extends an already existing algorithm for 2D detection on images to approximate 3D localization in space. A final probabilistic recursive estimation process combines the measures provided by the different sensors in order to improve the target pose computation, considering all the possible uncertainty and disturbance sources that may interfer, thus making the system more robust and efficient