Spatio-Temporal Trajectory Foundation Model: Recent Advances and Future Directions
Sean Bin Yang 1 , Ying Sun 2 , Yunyao Cheng 1 , Yan Lin 1 , Kristian Torp 1 , Jilin Hu 3 · 2025
Read Spatio-Temporal Trajectory Foundation Model: Recent Advances and Future Directions by Sean Bin Yang 1 , Ying Sun 2 , Yunyao Cheng 1 , Yan Lin 1 , Kristian Torp 1 , Jilin Hu 3 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-20729.txt.
Introduction
Sean Bin Yang and five colleagues present in 2025 a tutorial survey of trajectory foundation models, a young field adapting the foundation-model idea to movement data. A trajectory is a timed sequence of positions, the digital trail left by a person, vehicle, or animal. Researchers now want the same train-once flexibility for trajectory tasks like travel-time estimation and path ranking.
The tutorial runs as a 90-minute lecture in four parts, from basic definitions through data types and learning methods to open problems. It sorts methods by data type: raw GPS points, map-matched road segments, grid-based paths, and mixes adding text or images. Learning approaches range from contrastive methods, which teach models by comparing similar and dissimilar paths, to augmentation schemes that manufacture extra training views. The authors close with challenges around building models that are robust, responsible, and transferable, the systematic map the field lacked.
Catalog record
| CREATOR | Sean Bin Yang 1 , Ying Sun 2 , Yunyao Cheng 1 , Yan Lin 1 , Kristian Torp 1 , Jilin Hu 3 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-20729
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Spatio-Temporal Trajectory Foundation Model: Recent Advances and Future Directions Sean Bin Yang 1 , Ying Sun 2 , Yunyao Cheng 1 , Yan Lin 1 , Kristian Torp 1 , Jilin Hu 3 Abstract . Foundation models (FMs) have emerged as a powerful paradigm, enabling a diverse range of data analytics and knowledge discovery tasks across scientific fields. Inspired by the success of FMs—particularly large language models—researchers have recently begun to explore spatio-temporal foundation models (STFMs) to improve adaptability and generalization across a wide spectrum of spatio-temporal (ST) tasks. Despite rapid progress, a systematic investigation of trajectory foundation models (TFMs), a crucial subclass of STFMs, is largely lacking. This tutorial addresses this gap by offering a comprehensive overview of recent advances in TFMs, including a taxonomy of existing methodologies and a critical analysis of their strengths and limitations. In addition, the tutorial highlights open challenges and outlines promising research directions to advance spatio-temporal general intelligence through the development of robust, responsible, and transferable TFMs. 1. Introduction Self-supervised learning (SSL), such as contrastive learning ( Yang et al. 2021 ; Yang et al. 2022b ; Xu et al. 2025 ; Wei et al. 2025 ) , and generative learning ( Yang et al. 2023 ; Zhou et al. 2024 ; Chen et al. 2021 ; Lin et al. 2025b ; Lin et al. 2021 ) , has exhibited strong theoretical foundations and remarkable empirical performance across a wide range of domains, including natural language processing ( Devlin et al. 2019 ) , computer vision ( Tian et al. 2020 ) , and graph data ( You et al. 2020 ) . These SSL approaches enable models to learn high-quality representations from large volumes of unlabeled data, thereby reducing reliance on costly manual annotations and improving generalization in downstream tasks. Figure 1. The framework of TFMs for ST Trajectory data. As shown in Figure 1 , trajectory foundation models (TFMs)—also referred to in earlier literature as trajectory representation learning—serve as a fundamental enabler for a variety of intelligent transportation applications, including travel time estimation ( Yang et al. 2021 ; Yang et al. 2022b ; Xu et al. 2025 ; Wei et al. 2025 ; Yang et al. 2023 ; Zhou et al. 2024 ; Chen et al. 2021 ; Jiang et al. 2023 ; Lin et al. 2023 ; Waury et al. 2019 ) , traffic analysis ( Fang et al. 2021b ; Ma et al. 2024 ; Han et al. 2025 ; Zhao et al. 2025 ; Fang et al. 2022b ; Yang et al. 2022a ; Yang and Yang 2020 ; Jensen et al. 2024 ) , trajectory similarity computation ( Hu et al. 2024 ; Zhou et al. 2025b ; Wang et al. 2025 ; Ma et al. 2024 ; Chang et al. 2023 ; Deng et al. 2022 ; Fang et al. 2022a ; Zhou et al. 2025a ; Lin et al. 2024 ; Miao et al. 2025 ; Li et al. 2024a ) , trajectory recovery ( Lin et al. 2025a ; Wang et al. 2025 ) , trajectory clustering ( Wang et al. 2024 ; Fang et al. 2021a ; Yao et al. 2024 ; Si et al. 2025 ) , trajectory classification ( Yang et al. 2022b ; Liang et al. 2022 ; Zhou et al. 2024 ; Zhou et al. 2025b ; Xia et al. 2025 ; Luo et al. 2024 ) , and trajectory generation ( Wang et al. 2025 ; Zhu et al. 2024 ; Zhu et al. 2023 ) . In particular, a trajectory can be formally defined as T=\langle(x_{1},y_{1},t_{1}),(x_{2},y_{2},t_{2}),\cdots,(x_{n},y_{n},t_{n})\rangle , where each element represents a spatio-temporal point consisting of spatial coordinates (x_{i},y_{i}) and a corresponding timestamp t_{i} . Such trajectories inherently capture both spatial movement patterns and temporal dynamics, making them a rich yet challenging modality for effective representation learning. Tutorial Overview. With the rapid proliferation of spatio-temporal data, trajectory foundation models have emerged as a critical research direction, attracting increasing attention from both academia and industry. This tutorial provides a systematic and comprehensive overview of this evolving field. We begin by formalizing the core concepts of trajectory data mining and trajectory foundation models, establishing a unified basis for subsequent discussions. The objectives of this tutorial are threefold: (i) to elucidate the motivations for developing trajectory foundation models by analyzing the diverse modalities of trajectory data and their unique characteristics; (ii) to review and critically examine recently proposed approaches from a modality-oriented perspective, with particular emphasis on the challenges, limitations, and open problems associated with each modality; and (iii) to investigate the emerging paradigm of trajectory foundation model construction, presenting a structured taxonomy of learning methodologies and offering analytical insights into their theoretical foundations and practical implications. The tutorial is organized as follows: (1) Introduction, Motivation, and Foundation (10 mins) . (2) Data-Modality-Oriented Methods (20 mins) . (3) Foundation Model Learning Methods (50 mins) . (4) Conclusion and Future Directions (10 mins) . This tutorial features: (i) an extensive review of the literature from the past five years, encompassing both pioneering contributions and the latest state-of-the-art approaches, as developed across multiple research communities ranging from data management to machine learning; (ii) insights into recent advancements across diverse solution strategies, thereby enriching the research landscape with novel perspectives and methodological innovations. This tutorial will equip the audience with a fundamental comprehension of TFMs and inspire them to work in this dynamic and evolving area. Relation to Previous Tutorials. Recent tutorials related to spatio-temporal foundation models have primarily focused (i) on broader aspects of spatio-temporal data—such as time series, video, and trajectories—while overlooking several state-of-the-art methods specific to trajectory data ( Liang et al. 2025 ; Yang et al. 2025 ) ; (ii) on specific types of spatial-temporal data (such as time series ( Liang et al. 2024 ; Liu et al. 2025 ) ). In contrast, this tutorial offers a substantial expansion over prior efforts, providing a detailed exposition of recently proposed methodologies complemented by interactive exploration for the audience. Beyond introducing techniques for trajectory foundation models, the tutorial adopts a comprehensive perspective on the field, emphasizing the diversity of trajectory data modalities and the advancement of learning methods. A section introduces a novel taxonomy of TFMs, thereby opening up promising research avenues. To the best of our knowledge, this is the first tutorial that (i) presents an extensive coverage of the trajectory foundation model landscape; (ii) integrates the most recent developments in foundation models for spatio-temporal trajectory data; and (iii) outlines future directions informed by large-scale trajectory datasets and principles of responsible artificial intelligence (AI), with particular attention to sustainability-oriented AI. Audience and Expected Background. This tutorial is designed for researchers and data analysts, with a particular focus on recent advances in TFMs. It aims to foster new collaborations between members of the data management community and data science practitioners from diverse application domains, while simultaneously broadening interest in TFMs. The tutorial provides the necessary background to ensure that participants can follow the entire presentation, alongside in-depth technical discussions of state-of-the-art solutions. In addition, we critically examine the limitations of existing approaches and highlight key open research problems. The material is structured to be accessible to newcomers, yet sufficiently