Researchers Introduce M3-Former for Long-Term Vessel Trajectory Forecasting

A scientific preprint published on arXiv introduces M3-Former, a multimodal artificial intelligence framework designed for long-term vessel trajectory prediction. The architecture leverages pre-trained large language models (LLMs) to encode static ship attributes and navigational intent, aligning them with dynamic trajectory features.

To capture both high-level route planning and localized maneuvering, M3-Former incorporates a dual-granularity Mixture-of-Experts (MoE) structure alongside a Steering-Weighted Cross-Entropy loss designed to handle sparse turning samples in complex waterways.

Evaluated on real-world Danish AIS data, the model outperformed existing baselines across 1- to 4-hour horizons. In the 4-hour prediction task, M3-Former reduced Average Displacement Error (ADE) by 4.4% and Final Displacement Error (FDE) by 5.1% relative to the top benchmark.