WAVN: A Topology-Aware Visual Navigation Framework Fusing CNN and GNN

This is a collaborative visual homing framework for multi-robots in GPS-denied environments. It represents the environment as a topological graph using a hybrid CNN/GNN architecture, enabling a privacy-preserving solution for decentralized learning and relational reasoning.

Modern robotic systems increasingly rely on GPS for positioning and navigation, but in indoor spaces, underground areas, urban canyons, or hostile environments, GPS signals may be completely unavailable or severely attenuated. For robot teams that need to collaborate on tasks, how to achieve reliable visual navigation without global positioning is an urgent technical problem to solve.

Traditional visual navigation methods usually rely on the local perception of a single robot and struggle to leverage collective team-level knowledge. While centralized learning methods can integrate multi-robot data, they bring privacy risks and communication bottlenecks. The core innovation of WAVN lies in proposing a decentralized learning framework combined with topology-aware scene understanding, which not only protects data privacy but also enables knowledge sharing.

WAVN uses a unique hybrid architecture that models the environment as a topological graph:

• Image embeddings as nodes: Images captured at each location are processed through CNN feature extraction to become a node in the graph
• Navigation transitions as edges: The transition relationship when a robot moves from one location to another forms a directed edge in the graph
• Relational reasoning: GNN performs reasoning on the graph to learn topological relationships between locations and navigation strategies

The advantages of this representation are:

1. Topological abstraction: Abstracts complex visual environments into graph structures, reducing the complexity of navigation problems
2. Relational modeling: Explicitly models the reachability and transition relationships between locations
3. Scalability: New locations can be dynamically added as nodes, and new paths as edges
4. Privacy protection: Each robot maintains its own subgraph without sharing raw image data

The project uses EfficientNet B0 as a frozen feature extractor. This choice balances feature quality and computational efficiency:

• EfficientNet B0: A lightweight yet powerful CNN architecture suitable for edge deployment
• Replaceable design: Can be easily replaced with other backbone networks like ResNet or MobileNet
• Frozen weights: The CNN part remains frozen during training, focusing on training the GNN's graph reasoning ability

A key innovation of WAVN is the four-channel edge-enhanced graph design:

• RGB image channel: Provides standard visual features
• Edge segmentation image channel: Provides structured contour information

This dual-modal input allows the model to understand both the appearance features of the scene and its geometric structure, significantly improving navigation robustness in complex environments.

The GNN part uses a 2-layer Graph Convolutional Network (GCN):

• Graph convolution layer: Learns relationships between nodes and message passing
• Global pooling: Aggregates features of all nodes into a fixed-size graph representation
• Linear classification head: Maps the learned representation to the navigation decision space

The role of global pooling is to convert variable-length node sequences into fixed-dimensional vectors, which is crucial for subsequent navigation decisions.

The project code is well-organized and divided into several core modules:

This is the core module for graph construction, containing four key functions:

• get_feature_extractor: Defines the feature extraction backbone network, configures weights, and sets projection and target dimensions
• ExtractImageFeatures: Converts RGB and edge segmentation images into a 4-channel tensor
• resolve_dual_paths: Parses image pairs to ensure the integrity of pairing between current and destination locations
• BuildGlobalGraphFromCSV: Creates a global graph based on information from the aforementioned functions, defining nodes and edges