Context-Enriched Natural Language Descriptions of Vessel Trajectories

Kostas Patroumpas; Alexandros Troupiotis-Kapeliaris; Giannis Spiliopoulos; Panagiotis Betchavas; Dimitrios Skoutas; Dimitris Zissis; Nikos Bikakis

cs.AIMar 8, 2026

Context-Enriched Natural Language Descriptions of Vessel Trajectories

Authors: Kostas Patroumpas, Alexandros Troupiotis-Kapeliaris, Giannis Spiliopoulos, Panagiotis Betchavas, Dimitrios Skoutas, Dimitris Zissis, Nikos Bikakis

Organizations: Archimedes/Athena RC, Greece · University of the Aegean, Greece · Athena RC, Greece · Hellenic Mediterranean University, Greece

Abstract

We address the problem of transforming raw vessel trajectory data collected from AIS into structured and semantically enriched representations interpretable by humans and directly usable by machine reasoning systems. We propose a context-aware trajectory abstraction framework that segments noisy AIS sequences into distinct trips each consisting of clean, mobility-annotated episodes. Each episode is further enriched with multi-source contextual information, such as nearby geographic entities, offshore navigation features, and weather conditions. Crucially, such representations can support generation of controlled natural language descriptions using LLMs. We empirically examine the quality of such descriptions generated using several LLMs over AIS data along with open contextual features. By increasing semantic density and reducing spatiotemporal complexity, this abstraction can facilitate downstream analytics and enable integration with LLMs for higher-level maritime reasoning tasks.

Figures & tables

Fig. 1 : Originally reported AIS locations (in black dots) of a vessel after annotation: green boxes signify stops, orange triangles denote turning points, and red stars indicate noise.

Fig. 2 : Trip representation in successive episodes with statistics (in green box) along with contextual information from various external data sources (in red box) as extracted in CSV format suitable for map visualization (MAP).

Fig. 3 : Alternative episode representations.

Panel of Figure 4 — Fig. 4: Prompt template for generating trajectory descriptions. The LLM receives as input a json array with distilled context-enhanced representation about a vessel trip and generates: (i) semantically-rich natural language descriptions in plain text, and (ii) overall trip statistics (in json format ).

Panel of Figure 5 — Fig. 5 : Trajectory episodes (depicted on map with alternating colors) given as input to LLM for generating a textual description.

Fig. 6 : Distribution of annotated episodes among all trips.

Fig. 7 : Cumulative duration of annotated episodes in all trips.

Episode type	geospatial	weather	bythometry
SAILING	63	39	80
TURNING	54	37	98
MANEUVERING	41	12	64
STOPPED	100	0	37
COMMUNICATION GAP	0	0	0

TABLE I : Percentage (%) of episodes enriched with context.

LLM Model	response time (sec)
llama-3.1-8b-instant	1.01
llama-3.3-70b-versatile	1.12
qwen/qwen3-32b	2.58
openai/gpt-oss-20b	1.45
openai/gpt-oss-120b	2.36

TABLE II : Average response time in generating a trip description using different LLMs.

Panel of Figure 10 — Fig. 8 : Prompt template for evaluating the generated trajectory descriptions. The LLM receives as input a json array with distilled context-enhanced representation about a vessel trip and evaluates the quality of the generated NL description in terms of relevance, faithfulness, and correctness.

LLM Model	mean	stdev	max
llama-3.1-8b-instant	113.72	1047.73	19017.24
llama-3.3-70b-versatile	13.68	16.79	85.00
qwen/qwen3-32b	7.86	17.88	206.38
openai/gpt-oss-20b	1.87	6.74	88.10
openai/gpt-oss-120b	1.53	6.47	88.10

TABLE III : Statistics on deviations (%) of trip durations estimated by LLM models w.r.t. ground truth.

LLM Model	mean	stdev	max
llama-3.1-8b-instant	13.79	26.36	248.28
llama-3.3-70b-versatile	2.44	4.52	34.55
qwen/qwen3-32b	4.92	43.48	848.22
openai/gpt-oss-20b	1.06	4.29	35.32
openai/gpt-oss-120b	0.89	4.12	35.32

TABLE IV : Statistics on deviations (%) of travelled distance per trip as estimated by LLM models w.r.t. ground truth.

LLM Model	relevance	faithfulness	correctness
llama-3.1-8b-instant	3.90	3.84	3.27
llama-3.3-70b-versatile	4.47	4.77	4.36
qwen/qwen3-32b	4.42	4.48	3.82
openai/gpt-oss-20b	4.69	4.81	4.35
openai/gpt-oss-120b	4.91	4.96	4.69

TABLE V : Qualitative evaluation of textual descriptions of vessel trips generated by diverse LLM models.

Maritime Transportation

See all 19 papers

ShipTraj-R1: Reinforcing Ship Trajectory Prediction in Large Language Models via Group Relative Policy Optimization
Mar 3, 2026Yang Zhan, Yunhao Li, Zhang Chao +2Vessel Trajectory Prediction Trajectory Prediction
School of Artificial Intelligence, Optics, and Electronics (iOPEN), Northwestern Polytechnical University, Xi’an 710072, China · Department of Logistics and Maritime Studies, The Hong Kong Polytechnic University, Hong Kong 999077, China · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China
EnvShip-Bench: An Environment-Enhanced Benchmark for Short-Term Vessel Trajectory Prediction
Jun 13, 2026Kun Ma, Qilong Han, Chengjing Song +3Forecasting Benchmarks Vessel Trajectory Prediction
Harbin Engineering University Harbin, China · Politecnico di Torino Turin, Italy · Institute of Software, Chinese Academy of Sciences Beijing, China
M3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction
Sep 11, 2026Wenzhe Jin, Haina TangMixture of Experts Vessel Trajectory Prediction
School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China

Mar 3, 2026J/K move · Enter open · S save

ShipTraj-R1: Reinforcing Ship Trajectory Prediction in Large Language Models via Group Relative Policy Optimization

Yang Zhan, Yunhao Li, Zhang Chao, Yuxu Lu+1

School of Artificial Intelligence, Optics, and Electronics (iOPEN), Northwestern Polytechnical University, Xi’an 710072, China · Department of Logistics and Maritime Studies, The Hong Kong Polytechnic University, Hong Kong 999077, China · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China

Vessel Trajectory Prediction Trajectory Prediction Maritime Navigation Maritime Transportation

Open paper AskPDF arXiv

Recent advancements in reinforcement fine-tuning have significantly improved the reasoning ability of large language models (LLMs). In particular, methods such as group relative policy optimization (GRPO) have demonstrated strong capabilities across various fields. However, applying LLMs to ship trajectory prediction remains largely unexplored. In this paper, we propose ShipTraj-R1, a novel LLM-based framework that reformulates ship trajectory prediction as a text-to-text generation problem. (1) We design a dynamic prompt containing trajectory information about conflicting ships to guide the model to achieve adaptive chain-of-thought (CoT) reasoning. (2) We introduce a comprehensive rule-based reward mechanism to incentivize the reasoning format and prediction accuracy of the model. (3) Our ShipTraj-R1 is reinforced through the GRPO mechanism guided by domain-specific prompts and rewards, and utilizes the Qwen3 as the model backbone. Extensive experimental results on two complex and real-world maritime datasets show that the proposed ShipTraj-R1 achieves the least error compared with state-of-the-art deep learning and LLM-based baselines.

Abstract

Maritime Transportation

Explore similar work

ShipTraj-R1: Reinforcing Ship Trajectory Prediction in Large Language Models via Group Relative Policy Optimization

EnvShip-Bench: An Environment-Enhanced Benchmark for Short-Term Vessel Trajectory Prediction

M3-Former: Multimodal Transformer with Mixture-of-Experts for Long-Term Vessel Trajectory Prediction