AI Model Generates Realistic Daily Mobility Patterns for Urban Planning

Transportation planners and researchers have long struggled to model how people move through cities—a challenge that underpins everything from traffic management to pandemic response. Existing methods rely on either rigid assumptions and expensive surveys or deep learning approaches that miss important connections between activities and fail to adapt to new regions.

A New Generative Approach#

A team of researchers has proposed a Transformer-based generative model that takes a different path: rather than collecting exhaustive GPS traces or imposing inflexible rules, the system learns to synthesize realistic daily activity chains directly from socio-demographic and household attributes. The model begins by predicting sequences of activities (work, shopping, school dropoff, etc.), then uses a location module to assign those activities to specific geographic zones, yielding complete daily trajectories.

Training and Transfer Learning#

The approach is trained on openly available household travel survey data collected at the national level, then fine-tuned using local data. This dual-stage strategy allows the model to capture broad activity patterns while adapting to regional variation. The researchers demonstrated successful transfer to California, Washington, and Mexico City—suggesting the technique generalizes well across different urban contexts and data availability scenarios.

Real-World Validation#

The model was tested at scale through traffic simulations in Los Angeles County. When compared to the SCAG Activity-Based Model (a conventional benchmark), the new approach achieved location predictions with a cosine similarity of 0.997 and vehicle-miles-traveled estimates with a 4.97% mean absolute error. Against actual traffic observations from Caltrans PeMS sensors on California's I-405 corridor, the simulated traffic volumes matched real-world data within 5.85% error, with speed predictions accurate to 4.36%.

Implications#

The work addresses a genuine bottleneck in urban mobility research: the difficulty of modeling human behavior patterns in places where traditional data is scarce or outdated. By combining readily available surveys with generative AI, the method could equip planners with faster, cheaper tools for evaluating transportation infrastructure, testing policy interventions, and preparing for public health scenarios—all without waiting for new, costly data collection campaigns.