In modern, data-driven word, privacy becomes a significant concern for users of robotic systems. Federated learning (FL) is a machine learning paradigm in which a federation of clients is trained collaboratively, without sharing local datasets, consequently increasing their privacy. In this work, a novel heterogeneous FL framework is proposed, capable of training federations regardless of the model architectures used. With this framework, grasp prediction models are trained, and a pick-and-place pipeline is deployed, presenting the first application of FL in industrial robotics. In addition, the flexibility of the system is shown in image classification and sentiment analysis tasks. The influence of knowledge distillation on training results is also investigated.
As the results show, the presented FL framework can significantly improve client performance compared to training clients in isolation. At the same time, contrary to standard distributed learning approaches, it mitigates privacy risks introduced by data sharing.