Multi-Modal Fusion for End-to-End Underwater Navigation System

Underwater navigation poses unique challenges due to the inherent limitations of individual sensors. Cameras suffer from poor visibility while sonar provides only limited angular coverage. To overcome these limitations, a novel end-to-end navigation network is proposed that synergistically fuses data from multiple complementary sensors including camera, sonar, and positional information. The network employs an attention-based fusion mechanism to adaptively integrate multimodal features, directly generating control commands for autonomous navigation. This approach enables comprehensive environmental perception and precise motion control, particularly critical for maintaining safe proximity to the seabed. Extensive evaluations are performed in two different simulated underwater environments. The proposed network demonstrates advantages compared to traditional single-sensor approaches. Notably, in challenging scenarios requiring both obstacle avoidance and seabed cruising, the proposed method achieves an 80% success rate, providing an effectiveness enhancement for maintaining consistent trajectory stability.