GNN Federated Visual Homing: A New Breakthrough in Multi-Robot Collaborative Navigation Under GPS-Denied Environments

This project addresses the robot navigation challenges in GPS-denied environments (such as underground mines, indoor buildings, etc.). It proposes a hybrid architecture integrating CNN and GNN, achieves multi-robot collaborative visual homing via federated learning, and innovatively applies graph neural networks and distributed learning technologies, providing an effective solution for robot navigation in complex environments.

In GPS-denied scenarios, traditional navigation methods fail; visual homing is an alternative, but single robots face issues like limited field of view and difficulty in feature matching. Multi-robot collaboration can improve robustness, but it faces technical challenges in data privacy protection and communication coordination.

It uses CNN to extract visual features and generate high-dimensional embedding vectors. GNN constructs an environmental topological graph (image embeddings as nodes, navigation transition relationships as edges), realizes relational reasoning through message passing, and supports multi-hop long-distance navigation planning.

A federated learning framework is introduced: robots only upload model parameters during local training (to protect data privacy). An adaptive aggregation strategy is designed to solve the problem of data distribution differences, and differential privacy and gradient compression technologies are combined to reduce communication overhead.

In simulation and real-scenario tests, the collaborative scheme shows significant improvements in navigation success rate, path efficiency, and time consumption compared to single-robot visual homing. The federated learning version's performance is close to that of centralized training; increasing the number of robots enhances the model's generalization ability, and the performance of different GNN architectures is compared.

It can be applied in fields such as search and rescue (collapsed buildings/tunnels), agriculture (greenhouse/orchard patrol), and military (reconnaissance in denied environments). The federated learning framework can be extended to robot tasks like object recognition and semantic segmentation, providing a foundation for large-scale privacy protection systems.

Currently, it relies on general CNNs for feature extraction; navigation-specific visual encoders need to be optimized. GNNs have high computational costs for large-scale graphs, so efficient approximation algorithms need to be studied. In the future, SLAM technology will be integrated to handle dynamic environment changes, and code and datasets are planned to be open-sourced.

This project combines three technologies: deep learning, graph neural networks, and federated learning, to solve the multi-robot collaborative navigation problem in GPS-denied environments. It has academic value and practical application potential, demonstrating the application prospects of AI in complex scenarios.