Synthetic Human Mobility Generation from Urban Proxies

This direction is split into other directions: mobility data accessibility. Its last version stays readable and citable here.

Version 1 of 1117 papers123 conceptsfrozen 2026-10-04

Urban planning, epidemic forecasting, and traffic management all rely on a single, elusive ingredient: high-fidelity human mobility data. Historically, this data has been locked behind the "walled gardens" of telecommunication companies or restricted by stringent privacy regulations. Researchers face a "data desert" where the movement of millions is either too expensive to acquire or too sensitive to share.

To bypass these barriers, a new research direction focuses on synthetic mobility generation. Instead of collecting raw GPS pings, researchers are developing frameworks that "hallucinate" realistic human trajectories by leveraging public proxies—such as satellite imagery, social media footprints, and census data—coupled with generative AI.

The Physics of Movement as a Generative Prior

The intuition behind synthetic mobility is that human movement is not random; it is governed by physical and structural constraints. One of the most enduring concepts is the "Gravity Law," which posits that the flow between two locations is proportional to their "mass" (population or attractiveness) and inversely proportional to the distance between them.

Can Rong and Yong Li demonstrated this in ODGN, where they coupled the classical Gravity Law with Multi-view Graph Attention Networks. By using physics to guide a Generative Adversarial Network (GAN), the model can synthesize realistic Origin-Destination (OD) networks even for "unsurveyed" cities where no prior mobility data exists. This approach treats physics as a "sanity check" for the neural network, ensuring that generated flows respect the fundamental friction of distance.

Similarly, the GTG model developed by Jingyuan Wang and Ying Lin incorporates "Space Syntax" theory. By disentangling city-specific features from universal mobility patterns, they allow models to perform zero-shot generation, effectively "transferring" the logic of movement from a data-rich city like New York to a data-poor environment.

Visual and Social Proxies

If we cannot track the people, we can look at the stage they move upon. The "built environment"—roads, buildings, and parks—serves as a powerful proxy for human behavior.

Song Gao and Yichen Xu introduced Imagery2Flow, a deep learning framework that predicts neighborhood-to-neighborhood mobility flows using only medium-resolution satellite imagery. The model learns that certain visual features (like the density of commercial rooftops or the layout of transit hubs) are highly correlated with specific social and economic pulses.

Beyond the physical landscape, digital footprints offer a voluntary alternative to invasive tracking. Early work by James Caverlee and Zhiyuan Cheng on "Geo-Social Intelligence" showed that public social media check-ins could predict city rhythms. More recently, Jiajun Liu and Kun Zhao utilized geo-tagged tweets as a high-resolution proxy for population distribution, proving that even sparse digital signals can reconstruct the aggregate movement of a population.

Simulating Intent with Large Language Models

A significant limitation of traditional synthetic models is their lack of "reasoning." They can mimic the where and when, but they struggle with the why. This is where Large Language Models (LLMs) are being integrated to create "digital twins" of human agents.

The TrajLLM framework, proposed by Chen Ju and Hao Xue, uses LLMs to simulate the semantic intent behind a journey. By giving an agent a "memory" and a "personality," the model generates trajectories that are not just statistically plausible but logically consistent (e.g., an agent goes to a coffee shop because it is morning and they are "tired").

This shift toward agent-based modeling allows researchers like Hanqing Yang and Marie Siew to build AitL-RL, which optimizes complex systems like mall HVAC units by simulating how diverse groups of people would react to environmental changes, all without ever touching a single piece of private real-world data.

From Data Silos to Open Benchmarks

The ultimate goal of this research is to democratize urban science. By moving away from proprietary datasets, researchers are creating open-source alternatives that allow for reproducible science.

Projects like MOVEOD by Rishav Sen and Abhishek Dubey synthesize building-level commute data from public census records, while the YJMob100K dataset released by Kota Tsubouchi and Esteban Moro provides a standardized, anonymized benchmark for the community.

By synthesizing the "social fabric" of our cities from the ground up, these models ensure that urban intelligence is no longer a luxury reserved for data-rich corporations, but a tool available for public good.

Papers in this version

Showing 60 of 117 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_synthetic_human_mobility_generation_from_3c7a6e26,
  title = {Synthetic Human Mobility Generation from Urban Proxies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-human-mobility-generation-from-urban-proxies/rdr_3c7a6e2659bd4fbe9925693842782713},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_synthetic_human_mobility_generation_from_3c7a6e26_v1,
  title = {Synthetic Human Mobility Generation from Urban Proxies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/synthetic-human-mobility-generation-from-urban-proxies/rdr_3c7a6e2659bd4fbe9925693842782713/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}