Artificial Intelligence (AI) has proven to enable large economic efficiency gains in the context of the recent digital revolution in industry. Especially in the field of industrial robotics, AI has the potential to automate tedious, heavy, or complex tasks, and, thereby, assisting human workers. To function properly, the creation and training of AI-based systems typically needs large amounts of data. As many industrial manufacturing companies are small or medium sized, picking items are highly individual. This results in a small amount of data for a specific task. Sharing data across factories and companies is a promising approach toward gaining more data for industrial use cases. For example, the same robot model can be used for similar tasks at multiple factories and companies.

However, sharing data has proven to be challenging in real world applications, as companies do not like to share their critical production data with other companies or even competitors for fear of economic disadvantage when doing so. The resulting small datasets lead to less accurate AI models for robotic applications and the full potential of AI systems in industrial environments is not exploited. Federated learning is an emerging approach toward distributed, privacy-preserving machine learning. The training of AI-based systems is done locally and only AI model parameters are then uploaded to a central cloud server which is shared by multiple stakeholders. All stakeholders then benefit from an improved AI model, based on data collected at many different places without violating data privacy of the participating companies.

In this project, we aim to develop an international federated learning system in the domain of robotic picking and placing of unknown objects. The goal is to boost current AI solutions with more data while preserving privacy regulations. The usage of more versatile data allows us to create more feasible network structures and increase detection performance in comparison to single robot training. The AI models are designed to be as efficient as possible both to run locally in each side and globally on the central cloud server.