Distributed Deep Learning for Real-World Implicit Mapping in Multi-Robot Systems

Abstract

Using a distributed learning framework and 2D LiDAR data, this paper presents a novel approach for distributed navigation and environmental mapping. Each robot collects distance data in real time to create a local map, which is then shared and integrated among the robots through wireless communication. The framework uses DiNNO (Distributed Neural Network Optimization) for distributed learning to collaboratively optimize navigation paths and improve mapping accuracy. DiNNO effectively balances computational load and communication overhead while providing superior accuracy and efficiency compared to other algorithms. Continuous data collection and real-time map updates ensure adaptability to dynamic environments. Experimental results demonstrate the effectiveness of the system and highlight its potential for a variety of autonomous navigation and mapping applications. This approach, as enhanced by DiNNO, offers significant advantages in terms of communication efficiency and mapping accuracy, thus providing a robust solution for dynamic and complex environments.