The expertise of German companies in the field of production technology and mechanical engineering is an important toothed wheel in Europe's strongest economy. The FEATHER project uses these successes and investigates possibilities for a uniform, data protection-compliant modernization of industrial robotics. By using federated learning, we want to turn complex production robot interactions into a challenge that transcends company boundaries in order to solve these problems cooperatively.

A major problem with machine learning in industrial scenarios is the shortage of high-quality data. The federated learning paradigm addresses this challenge by companies cooperatively improving their algorithms. This is done without publishing internal company data, so internal processes and secrets are secured. One of FEATHER's goals is to develop a standardized platform that enables production companies to cooperate in  this way.

The scenario is examined in more detail using the example of so-called soft end effectors. Soft end effectors are flexible tool holders for robot grippers which, due to their nature, offer greater task adaptability compared to rigid holders. The planning of motion sequences is difficult to calculate due to the flexible material properties of such end effectors. Furthermore, the use of such tool holders is more of a rarity than the norm. This provides a perfect evaluation scenario for making progress in the field of physical robotics on the one hand and progress in the field of federated learning on the other.

The unified platform for federated learning should integrate modern computational paradigms such as cloud, edge and fog to meet the requirements. In addition, standardized approaches for the communication infrastructure and the organization of distributed resources will be integrated. This integration makes it possible for even companies with different IT infrastructures to participate seamlessly in federated learning processes. By using edge and fog computing, data processing is moved closer to where the data is generated, which reduces latency and increases efficiency. The cloud component ensures scalability and central coordination.

Highly adaptable robot systems are essential for the future of industry. Instead of rigid, specialized robots, we need systems that can adapt autonomously to new products and processes. By integrating federated learning, more robust control algorithms can be developed that enable robots to be integrated into diverse production scenarios more quickly and efficiently.

Innovative communication and orchestration technologies are essential for the success of production-oriented federated learning. The focus is on integrating edge and fog computing paradigms to move data processing closer to the point of origin, reducing latency and increasing efficiency. Standardized communication protocols and orchestration tools ensure seamless interaction between heterogeneous machines and systems.

In order to successfully implement federated learning in production, important empirical values must be taken into account. The heterogeneity of data sources requires robust aggregation mechanisms, and the platform must be scalable to manage a large number of decentralized endpoints. Data sovereignty is also crucial: companies must be able to trust that their sensitive production data will remain protected and that only necessary model information will be shared.

Ongoing project 01.06.2025 - 31.05.2028

SCHUNK SE & Co. KG - https://schunk.com/de/de

SEW-EURODRIVE GmbH & Co. KG -https://www.sew-eurodrive.de/startseite.html

VisionTools Bildanalyse Systeme GmbH - https://www.vision-tools.com/