TrajPrism: A Multi-Task Benchmark for Language-Grounded Urban Trajectory Understanding

Researchers introduced TrajPrism, a comprehensive benchmark dataset combining 300K real urban trajectories with natural language annotations across three cities, enabling AI models to understand the alignment between physical travel paths and human descriptions of movement intent, constraints, and preferences.

TrajPrism addresses a fundamental gap in AI research by bridging trajectory modeling and natural language understanding in urban mobility contexts. Historically, these domains have developed separately—trajectory research focused on geometric patterns while language-centric mobility work emphasized route planning and navigation tools rather than fine-grained text-trajectory alignment. This benchmark unifies three complementary tasks: generating trajectories from natural language instructions, retrieving trajectories matching semantic queries, and generating captions describing trajectories, creating a more holistic evaluation framework.

The dataset's construction across Porto, San Francisco, and Beijing with judge-filtered annotations under a four-dimensional travel-intent taxonomy reflects growing recognition that modern transportation systems must interpret human intent beyond simple A-to-B routing. With 2.1M task instances, TrajPrism provides sufficient scale for training production-grade models. The researchers' proof-of-concept models—TrajAnchor, TrajFuse, and TrajRap—demonstrate that geometry-only baselines significantly underperform when language grounds the task, validating the benchmark's design philosophy.

For the AI and urban computing communities, this work enables development of more sophisticated mobility AI systems that understand nuanced travel preferences expressed in natural language. The portable annotation pipeline designed for different cities suggests potential for expanding the benchmark and accelerating research in language-grounded transportation AI. As autonomous vehicles and mobility services increasingly rely on natural language interfaces for user interaction, benchmarks like TrajPrism become foundational infrastructure. The release with reproducible methodology allows other research groups to extend coverage to additional cities, potentially creating a comprehensive multilingual, multicultural resource for trajectory-language research.

• →TrajPrism combines 300K urban trajectories with language annotations across three major cities, creating 2.1M task instances for evaluating language-trajectory alignment.
• →The benchmark unifies three AI tasks: instruction-conditioned trajectory generation, semantic trajectory retrieval, and trajectory captioning with unified evaluation protocols.
• →Proof-of-concept models reveal substantial performance gaps between geometry-only baselines and language-aware approaches, validating the benchmark's design.
• →The portable annotation pipeline is designed for expansion across additional cities with compatible trajectory data and map resources.
• →This work addresses a research gap where trajectory modeling and natural language understanding have historically developed separately in mobility contexts.