TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models

Peiran Li; Jiawei Wang; Haoran Zhang; Xiaodan Shi; Noboru Koshizuka; Chihiro Shimizu; Renhe Jiang

cs.LGMar 16, 2026

TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models

Authors: Peiran Li, Jiawei Wang, Haoran Zhang, Xiaodan Shi, Noboru Koshizuka, Chihiro Shimizu, Renhe Jiang

Organizations: Hitotsubashi University · The University of Tokyo · LocationMind Inc. · Stockholm University

Abstract

The importance of mobile phone GPS trajectory data is widely recognized across many fields, yet the use of real data is often hindered by privacy concerns, limited accessibility, and high acquisition costs. As a result, generating pseudo-GPS trajectory data has become an active area of research. Recent diffusion-based approaches have achieved strong fidelity but remain limited in spatial scale (small urban areas), transportation-mode diversity, and efficiency (requiring numerous sampling steps). To address these challenges, we introduce TrajFlow, which to the best of our knowledge is the first flow-matching-based generative model for GPS trajectory generation. TrajFlow leverages the flow-matching paradigm to improve robustness and efficiency across multiple geospatial scales, and incorporates a trajectory harmonization and reconstruction strategy to jointly address scalability, diversity, and efficiency. Using a nationwide mobile phone GPS dataset with millions of trajectories across Japan, we show that TrajFlow or its variants consistently outperform diffusion-based and deep generative baselines at urban, metropolitan, and nationwide levels. As the first nationwide, multi-scale GPS trajectory generation model, TrajFlow demonstrates strong potential to support inter-region urban planning, traffic management, and disaster response, thereby advancing the resilience and intelligence of future mobility systems.

Figures & tables

Figure 1: In pseudo-human mobility generation, three key challenges remain to be addressed: multi-scale capability, transportation-mode diversity, and training & inference efficiency.

Panel of Figure 2 — Figure 2: (a) shows the accuracy degradation of DiffTraj at increasing spatial scales. (b) shows the SNR collapses when applying a fixed noising parameter across scales.

Figure 3: The overview of the proposed TrajFlow.

Central Tokyo
Method	Density JS $\downarrow$	DTW ${}_{\text{med}}$ $\downarrow$	Fr ${}_{\text{med}}$ $\downarrow$	DTW ${}_{\text{IQR}}$ $\downarrow$	DTW ${}_{\text{P10}}$ $\downarrow$	DTW ${}_{\text{P90}}$ $\downarrow$	Fr ${}_{\text{IQR}}$ $\downarrow$	Fr ${}_{\text{P10}}$ $\downarrow$	Fr ${}_{\text{P90}}$ $\downarrow$
TrajFlow (ours)	0.0674	20.350	0.304	13.392	10.574	39.119	0.174	0.200	0.674
TrajFlow-w/o-OD	0.3560	916.436	13.862	708.183	432.567	1,740.890	7.313	7.064	20.858
TrajFlow-w/o-RDP	0.0642	22.088	0.340	16.491	11.149	47.959	0.238	0.209	0.873
TrajFlow-w/o RDP & OD	0.0323	8.179	0.184	4.586	4.994	14.159	0.118	0.110	0.363
DiffTraj	0.1340	44.321	0.651	40.713	19.399	109.349	0.544	0.341	1.774

Table 1: Evaluation grouped by region (unit = km). Best metric in bold .

Figure 4: Visualization of trajectory samples. Ground-truth and generated nationwide trajectories are shown with zoomed views highlighting three representative regions: (A) Tokyo Metropolis, (B) Tohoku Area, and (C) Kansai Area, across all generative models.

Figure 5: Per-mode average trip distance in Tokyo: ground truth vs. generated.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Param 5		Param 10		Param 20		Param 30
Method	DTW	Fréchet	DTW	Fréchet	DTW	Fréchet	DTW	Fréchet
direct_k	3.78	0.0908	1.54	0.0430	0.79	0.0205	0.66	0.0155
dct	3.23	0.0814	1.24	0.0326	0.69	0.0150	0.61	0.0137
rdp_k	2.77	0.0580	0.95	0.0219	0.61	0.0143	0.59	0.0150
anchor	16.42	0.3333	1.75	0.0405	0.76	0.0191	0.61	0.0138
spline_lsq	3.11	0.0665	1.22	0.0271	0.68	0.0150	0.61	0.0136

Appendix

Table 2: Reconstruction Performance (Avg DTW/Fréchet, unit=km). Best (lowest) values in bold .

Central Tokyo
Method	Density JS $\downarrow$	DTW ${}_{\text{med}}$ $\downarrow$	Fr ${}_{\text{med}}$ $\downarrow$	DTW ${}_{\text{IQR}}$ $\downarrow$	DTW ${}_{\text{P10}}$ $\downarrow$	DTW ${}_{\text{P90}}$ $\downarrow$	Fr ${}_{\text{IQR}}$ $\downarrow$	Fr ${}_{\text{P10}}$ $\downarrow$	Fr ${}_{\text{P90}}$ $\downarrow$
TrajFlow (ours)	0.0674	20.350	0.304	13.392	10.574	39.119	0.174	0.200	0.674
TrajFlow-w/o-FM (steps=10)	0.3213	362.611	5.844	794.886	86.305	2096.804	14.045	1.149	37.590
TrajFlow-w/o-FM (steps=50)	0.2923	263.889	4.390	694.411	64.512	1827.793	12.156	0.855	32.903
TrajFlow-w/o-FM (steps=100)	0.2269	142.836	2.363	303.280	38.252	749.346	5.244	0.519	12.811
TrajFlow-w/o-FM (steps=200)	0.1208	36.434	0.583	49.489	15.035	130.881	0.975	0.247	2.437

Appendix

Table 3: TrajFlow-w/o-FM performance with different training/sampling steps (unit: km). Best metric in bold .

Figure 10: Bar plot for time cost comparison in inference stage.

Region	Method	Steps	Sec/Sample	Sampling Sec	Total Sec
Central Tokyo	TrajFlow	10	0.0097	0.9696	4.8912
Central Tokyo	TrajFlow-w/o-FM (steps=10)	10	0.0100	1.0039	4.9646
Central Tokyo	TrajFlow-w/o-FM (steps=50)	50	0.0382	3.8235	7.7105
Central Tokyo	TrajFlow-w/o-FM (steps=100)	100	0.0761	7.6105	11.5326
Central Tokyo	TrajFlow-w/o-FM (steps=200)	200	0.1585	15.8546	19.8442
Central Tokyo	TrajFlow-w/o-FM (steps=300)	300	0.2237	22.3682	26.3010

Appendix

Table 4: Comparison table for time cost in inference stage.

Panel of Figure 11 — Figure 11: Transportation-mode diversity visualization: GT vs. Gen in Tokyo.

Panel of Figure 12 — Figure 12: The capability to generate GPS trajectories under given transportation mode conditions. Ground-truth nation-wide trajectories shown with zoomed views highlighting three representative regions: (A) Tokyo Metropolis, (B) Tokaido Area, and (C) Kansai Area, across various transportation modes: Train, Car, Walk, Bike (a) shows the ground-truth trajectories. (b) shows the generated.

uid	segment_id	trans_mode1	trans_mode2	time	lat	lon
…	1	STAY	STAY	2023/2/6 0:00	35.9497	139.5576
…	1	STAY	STAY	2023/2/6 0:15	35.9497	139.5576
…	…	…	…	…	…	…
…	2	MOVE	TRAIN	2023/2/6 0:36	36.2269	114.1721
…	…	…	…	…	…	…
…	3	MOVE	WALK	2023/2/6 0:59	36.7236	114.4598

Appendix

Table 5: Example rows of the dataset we used. (Note: this example data is only for showing data format without real lat/lon values.)

Chengdu
Method	Density JS $\downarrow$	DTW ${}_{\text{med}}$ $\downarrow$	Fr ${}_{\text{med}}$ $\downarrow$	DTW ${}_{\text{IQR}}$ $\downarrow$	DTW ${}_{\text{P10}}$ $\downarrow$	DTW ${}_{\text{P90}}$ $\downarrow$	Fr ${}_{\text{IQR}}$ $\downarrow$	Fr ${}_{\text{P10}}$ $\downarrow$	Fr ${}_{\text{P90}}$ $\downarrow$
TrajFlow (ours)	0.211	35.911	0.547	29.418	15.839	72.048	0.392	0.279	0.959
TrajFlow-w/o-OD	0.0809	3.696	0.123	2.847	2.180	8.904	0.117	0.0614	0.362
TrajFlow-w/o RDP	0.206	34.687	0.540	30.020	14.650	69.999	0.383	0.266	0.957
TrajFlow-w/o RDP & OD	0.0140	2.158	0.0525	0.896	1.519	3.278	0.0243	0.0354	0.0880
TrajFlow-w/o Flow	0.208	36.457	0.561	30.017	15.948	73.523	0.397	0.278	0.975

Appendix

Table 6: Evaluation grouped by region (unit = km). Best metric in bold .

Trajectory Generation

Trajectory generation creates plausible, goal-directed paths and motion sequences for people, vehicles, and robots.

See all 182 papers

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation
May 11, 2026Wilson Wongso, Lihuan Li, Arian Prabowo +4Trajectory Generation Synthetic Data Generation
University of New South Wales · Hong Kong University of Science and Technology (Guangzhou)
From GPS Points to Travel Patterns: Flexible and Semantic Trajectory Generation with LLMs
May 28, 2026Silin Zhou, Chenhao Wang, Yuntao Wen +3Trajectory Generation Intelligent Transportation Systems
University of Electronic Science and Technology of China Chengdu, China · King Abdullah University of Science and Technology Thuwal, Saudi Arabia
Plan-to-Synthesis: Cross-City Human Mobility Generation via Semantic Latent Flow Matching
Sep 26, 2026Zhoufu Wang, Baoshen Guo, Zhiqing Hong +6Trajectory Generation Generative Modeling
Singapore-MIT Alliance for Research and Technology Centre, MIT · Nanyang Technological University · Hong Kong University of Science and Technology (Guangzhou) +3

May 11, 2026J/K move · Enter open · S save

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation

Wilson Wongso, Lihuan Li, Arian Prabowo, Xiachong Lin+3

University of New South Wales · Hong Kong University of Science and Technology (Guangzhou)

Trajectory Generation Synthetic Data Generation Diffusion Language Models Diffusion Models

Open paper AskPDF arXiv

Generating high-fidelity synthetic GPS trajectories is increasingly important for applications in transportation, urban planning, and what-if scenario simulation, especially as privacy concerns limit access to real-world mobility data. Existing trajectory generation models face a trade-off between efficiency and faithfulness to road network topology: continuous-space methods enable fast generation but ignore the road network, while topology-aware approaches rely on search-based autoregressive decoding that limits generation speed. We propose TrajDLM, a topology-aware trajectory generation framework based on block diffusion language models that bridges this gap. TrajDLM models trajectories as sequences of discrete road segments, combining a block diffusion backbone for efficient denoising, topology-aware embeddings from a road network encoder, and topology-constrained sampling to ensure coherent and realistic trajectories. Across three city-scale datasets, TrajDLM achieves strong performance on fine-grained local similarity metrics while being up to

2.8\times

faster than prior work, and demonstrates strong zero-shot transfer across domains, including unseen transportation modes. These results highlight the effectiveness of block-wise discrete diffusion as a scalable approach to accurate and efficient trajectory generation. Our code is available at https://github.com/cruiseresearchgroup/TrajDLM/

Abstract

Trajectory Generation

Explore similar work

TrajDLM: Topology-Aware Block Diffusion Language Model for Trajectory Generation

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

Plan-to-Synthesis: Cross-City Human Mobility Generation via Semantic Latent Flow Matching