Real-Time Data Processing Architecture for Multi-Robots Based on Differential Federated Learning

Abstract

The emergency of ubiquitous intelligence in various things has become the ultimate cornerstone in building a smart interconnection of the physical world and the human world, which also caters to the idea of Internet of Things (IoT). Nowadays, robots as a new type of ubiquitous IoT devices have gained much attention. With the increasing number of distributed multi-robots, such smart environment generates unprecedented amounts of data. Robotic applications are faced with challenges of such big data: the serious real-time assurance and data privacy. Therefore, in order to obtain the big data values via knowledge sharing under the premise of ensuring the real-time data processing and data privacy, we propose a real-time data processing architecture for multi-robots based on the differential federated learning, called RT-robots architecture. A global shared model with differential privacy protection is trained on the cloud iteratively and distributed to multiple edge robots in each round, and the robotic tasks are processed locally in real time. Our implementation and experiments demonstrate that our architecture can be applied on multiple robotic recognition tasks, balance the trade-off between the performance and privacy.