A Hybrid Federated Reinforcement Learning Approach for Networked Robots

Federated learning (FL) evolved as a game changer to ensure the privacy of user-sensitive data, by using locally available models, rather than data, for aggregation at a central server. While in a centralized FL approach, network connectivity or server failures could bring down the system, its decentralized counterpart can circumvent this issue, but at the cost of increased learning times. With FL being used in the domain of networked robotics, a combination of centralized and decentralized approaches can prove to be a safer and more viable option. In this paper, we present a mobile agent-based Hybrid version of Federated Reinforcement Learning (HyFRL), where the learned models, viz. Q-tables, are aggregated and shared among a set of connected robots inhabiting different environments. Multi-robot experiments performed using several networked instantiations of Webots®, an open-source robot simulator, reveal the efficacy of this hybrid version over its centralized and decentralized equivalents.