Festo’s Federated Learning Enhances Robot Picking Efficiency and Protects Sensitive Data
German automation company Festo, in collaboration with the Karlsruher Institut für Technologie (KIT), the University of Waterloo, and Darwin AI, recently concluded its FLAIROP (Federated Learning for Robot Picking) research project. The two-year project aimed to enhance the intelligence of picking robots using distributed artificial intelligence (AI) methods while also protecting sensitive data.
The key focus of the research was to enable robots to learn from each other without the need to share their training data. This approach, known as Federated Learning, allows for the development of more robust and efficient AI models by utilizing data from multiple robots, all while safeguarding sensitive company information.
Jan Seyler, Head of Advanced Development Analytics and Control at Festo, expressed pride in the project’s success, stating, We have succeeded in showing that robots can learn from each other without sharing sensitive data and company secrets. This protects our customers’ data, and we also gain speed because the robots can take over many tasks faster this way.
By leveraging Federated Learning, Festo and its partners were able to create a universal, simulation-based dataset that enables autonomous gripping robots to reliably grasp unseen items. This breakthrough has the potential to significantly increase the efficiency and accuracy of robot picking operations.
Moving forward, the Federated Learning System will be further developed to allow different companies to train robot systems together without data sharing. This advancement could lead to increased acceptance and utilization of such systems in practical applications.
Federated Learning is a machine learning technique that prioritizes privacy-preserving AI applications. Unlike traditional methods where training data is sent to a central server, Federated Learning allows training to occur at various locations. Locally trained models are then aggregated on a central machine learning server, ensuring the sensitive training data remains with the respective data provider.
In the FLAIROP project, robot arms equipped with cameras were used to visually detect items in picking cells. Based on the camera images, the robots could recognize different items and select appropriate gripping methods. The challenge of dealing with the diverse range of items found in a warehouse was overcome by collecting large amounts of data from picking cells in different organizations.
Throughout the project, five autonomous picking stations were set up for training the robots, with two at the KIT Institute for Materials Handling and Logistics Systems and three at Festo SE & Co. KG. The final event showcased the usability of the research results, demonstrating how these advancements can be incorporated into Festo products.
The research findings will be published and made freely available for interested parties to utilize in their initial pilot projects. By sharing the knowledge gained from the FLAIROP project, Festo aims to foster innovation and collaboration in the field of robotic picking.
Overall, Festo’s Federated Learning research project has shown great promise in enhancing robot picking efficiency while protecting sensitive data. The use of distributed AI methods and the development of a privacy-preserving learning system pave the way for more advanced and secure automation in various industries.