# MarineFormer: A Spatio-Temporal Attention Model for USV Navigation in Dynamic Marine Environments

**Authors:** Ehsan Kazemi, Dechen Gao, Iman Soltani  
**arXiv:** [2410.13973](https://arxiv.org/abs/2410.13973)  
**Date:** 2024-10-17  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2410.13973v4/

## Key points

- Challenges in autonomous navigation in marine environments include spatially varying flow disturbances and dynamic obstacles
- Study by Ehsan Kazemi, Dechen Gao, and Iman Soltani introduces innovative solution leveraging local flow field measurements
- Integration of flow data with traditional sensory inputs like ego-state and obstacle states is key insight
- Proposal of MarineFormer policy architecture based on Transformer models with spatial and temporal attention mechanisms
- End-to-end training using reinforcement learning in a 2D simulated environment shows remarkable performance improvements
- MarineFormer approach enhances episode completion success rates by nearly 23% and reduces path length compared to classical and state-of-the-art baselines
- Ablation studies highlight the importance of flow measurements and effectiveness of proposed architecture
- Represents significant advancement in autonomous navigation systems for unmanned surface vehicles (USVs) operating in dynamic marine environments

## Summary

In the realm of autonomous navigation in marine environments, the challenges posed by spatially varying flow disturbances and dynamic obstacles are significant. However, a recent study by Ehsan Kazemi, Dechen Gao, and Iman Soltani introduces an innovative solution that leverages local flow field measurements to transform previously unsolvable navigation scenarios into manageable ones. The key insight lies in effectively integrating flow data with traditional sensory inputs like ego-state and obstacle states. The researchers propose a novel policy architecture called MarineFormer, which is based on Transformer models. This architecture incorporates two crucial attention mechanisms: spatial attention for sensor fusion and temporal attention for capturing environmental dynamics. Through end-to-end training using reinforcement learning in a 2D simulated environment with realistic flow features and obstacles, MarineFormer demonstrates remarkable performance improvements. Comparative evaluations against classical and state-of-the-art baselines reveal that the MarineFormer approach enhances episode completion success rates by nearly 23% while also reducing path length. Ablation studies further underscore the essential role of flow measurements and validate the effectiveness of the proposed architecture in effectively leveraging this data. Overall, this research represents a significant advancement in autonomous navigation systems for unmanned surface vehicles (USVs) operating in dynamic marine environments. By integrating spatio-temporal attention mechanisms with flow field measurements, MarineFormer offers a promising solution to address the complexities associated with navigating through challenging maritime conditions.