Over the past two years,Festo has been conducting research together with the Karlsruhe Institute of Technology (KIT) and partners from Canada (University of Waterloo, Darwin AI) to make picking robots more intelligent using distributed AI methods. To this end, the partners in the Flairop (Federated Learning for Robot Picking) project investigated how the most versatile training data possible from several plants or even companies can be used to develop more robust and efficient AI algorithms than with data from just one robot - without having to disclose sensitive company data.

"We are pleased that we have been able to show that robots can learn from each other without sharing sensitive data and trade secrets. This allows us to protect our customers' data and we also gain speed because the robots can take on many tasks more quickly in this way. For example, the collaborative robots can support production workers with repetitive, heavy and tiring tasks," says Jan Seyler, Head of Advanced Development Analytics and Control at Festo.

"We have developed a universal, simulation-based data set that we can use to train autonomous gripper robots so that they are able to reliably grip items that they have never seen before," explains Maximilian Gilles from KIT.

In the future, the federated learning system is to be further developed so that the platform enables different companies to train robot systems together without having to share data with each other.

Federated learning is a machine learning technique used to create privacy-preserving AI applications. Instead of sending the training data of the robot arms in the picking cells to a central server to train the model there, the training takes place at many different locations. The locally trained models are then sent to the central server for machine learning so that the sensitive training data does not leave the data provider. Nevertheless, federated learning enables learning across data silos by aggregating the distributed models and ultimately enabling highly accurate and data-driven prediction of object recognition and grasp points.

The robot arms in the picking cells are equipped with cameras to visually detect the objects in front of them. Based on the camera image, the robot arms automatically recognize the different items and select a suitable gripping method. Due to the variety of items in an industrial warehouse, this is a complicated task and large amounts of data are required to achieve reasonable results. Creating such large amounts of data is time consuming. With data collected from picking cells in different organizations, it was possible to improve the grasp prediction for picking cells.

During the project, a total of five autonomous picking stations were set up to train the robots: two at the KIT Institute for Materials Handling and Logistics Systems (IFL) in Karlsruhe and three at Festo, based in Esslingen am Neckar.
At the final event, Festo focused on the usability of the results. "We show which Festo products it can be used in. The research results will now be published and can be used freely by all interested parties in initial pilot projects," says Jan Seyler.