Flow-FL: Data-Driven Federated Learning for Spatio-Temporal Predictions in Multi-Robot Systems

In this paper, we show how the Federated Learning (FL) framework enables\nlearning collectively from distributed data in connected robot teams. This\nframework typically works with clients collecting data locally, updating neural\nnetwork weights of their model, and sending updates to a server for aggregation\ninto a global model. We explore the design space of FL by comparing two\nvariants of this concept. The first variant follows the traditional FL approach\nin which a server aggregates the local models. In the second variant, that we\ncall Flow-FL, the aggregation process is serverless thanks to the use of a\ngossip-based shared data structure. In both variants, we use a data-driven\nmechanism to synchronize the learning process in which robots contribute model\nupdates when they collect sufficient data. We validate our approach with an\nagent trajectory forecasting problem in a multi-agent setting. Using a\ncentralized implementation as a baseline, we study the effects of staggered\nonline data collection, and variations in data flow, number of participating\nrobots, and time delays introduced by the decentralization of the framework in\na multi-robot setting.\n