Federated Reinforcement Learning Framework for Mobile Robot Navigation Using ROS and Gazebo

Abstract

Mobile robot navigation involves guiding a robot from one location to another, a complex task requiring the integration of various components. Reinforcement learning (RL) has emerged as a promising approach for adaptive and efficient navigation without the need for pre-mapping. However, RL-based models often suffer from poor generalization in highly variant environments. Federated reinforcement learning (FRL), which combines federated learning (FL) and RL, presents a potential solution by improving model generalization while maintaining data privacy. Despite its advantages, the application of FRL in robotics remains largely unexplored, with a lack of dedicated frameworks to validate its effectiveness. In this paper, we propose an FRL framework for mobile robot navigation, using the robot operating system (ROS) and the Gazebo simulator. In our framework, RL is conducted on ROS1 machines acting as FRL clients, while model aggregation is managed on a ROS2-based server. We evaluate our framework in various unseen environments, demonstrating improved navigation performance and generalization capabilities compared to non-federated RL models.