From GPS Points to Travel Patterns: Flexible and Semantic Trajectory Generation with LLMs

Researchers propose HTP, an LLM-based framework that generates realistic urban trajectories by first synthesizing travel patterns and then GPS points, addressing privacy concerns in smart city applications. The method outperforms existing approaches by 29.78% and can generate variable-length trajectories under multiple conditions, advancing synthetic data generation for urban analytics.

This research tackles a critical infrastructure challenge: how to generate realistic trajectory data for smart city applications without exposing sensitive location information. Traditional trajectory datasets pose significant privacy risks, limiting their availability for researchers and developers building urban analytics systems. HTP addresses this by introducing a hierarchical generation approach that mirrors how humans conceptualize travel—first understanding the pattern or intent, then the specific route details. The methodology combines residual quantization variational autoencoders with large language models, treating trajectory generation as a semantic task rather than pure point prediction. This represents a meaningful shift in how synthetic mobility data can be created.

The broader context involves increasing regulatory pressure around location data privacy, particularly with GDPR and similar frameworks making large-scale trajectory collection legally risky. Synthetic trajectory generation offers an elegant solution for researchers, urban planners, and smart city developers who need realistic data without privacy violations. The 29.78% improvement over baselines suggests meaningful advancement in generation quality, which directly impacts the utility of synthetic datasets for model training and urban simulation.

For practitioners in smart cities, autonomous vehicles, and location analytics, this work enables safer experimentation with realistic movement patterns. Developers can now train systems on synthetic data that captures nuanced travel behaviors—such as traffic-induced density variations—without legal or ethical complications. The variable-length generation capability particularly matters, as real trajectories vary significantly in duration and complexity. The open-source availability accelerates adoption across academic and commercial sectors working on urban intelligence.

• →HTP hierarchically generates travel patterns before GPS points, better capturing realistic urban movement semantics
• →The method achieves 29.78% improvement over baseline approaches in generation quality metrics
• →Synthetic trajectory generation enables privacy-safe data access for smart city applications and urban analytics
• →LLM-based trajectory synthesis can handle variable-length sequences under multiple conditional constraints
• →Open-source code release accelerates adoption in urban planning, autonomous systems, and mobility research