Handling robot

A neural network determines the position and orientation of each part in the camera image, which is turned into gripping commands for the robot. In the urobots solution, the AI runs directly on the FPGA of the IDS NXT camera. The software converts 2D image coordinates into 3D coordinates for the robot. According to urobots, it captures the positions of all objects in around 200 milliseconds, which means five captures per second.

Not for training the AI: customers upload suitable images to the IDS NXT lighthouse cloud software, label them and start training at the push of a button. A VisionApp from urobots controls the robot; its software is written in Python, while the camera app uses C++. Because the neural network runs directly on the camera, the PC that would otherwise be required is no longer needed. Thanks to Ethernet, the app can be adapted to other robot models.

With a small number of labelled images and the knowledge of a subject expert. For inspecting sealing rings, 10 to 20 images per object were sufficient. An expert able to judge product defects sorted the images into good and defective. urobots trained the network, converted it to the IDS NXT format and uploaded it to the camera. The software then turns detections into work instructions that the robot uses to sort out defective rings.

With a digital service portal that brings together workflows, checklists and equipment history. Instead of Excel spreadsheets, Fabmatics uses the EquipmentCloud from Kontron AIS. The Workflows app guides technicians through tasks such as the precise calibration of cameras. Fabmatics reduced the time needed for commissioning by 50 percent and increased overall process speed by 30 percent; on average, transport robots now need only one acceptance test.