AI Agents Now Mimic Human Mobility Realistically

Understanding and predicting human mobility is crucial for designing smarter cities, from optimizing traffic flows to planning emergency evacuations. However, real trajectory data is often sparse and raises privacy concerns, limiting its use. Researchers have developed a novel AI framework that generates synthetic human mobility data by simulating the cognitive processes behind daily decisions, providing a scalable and interpretable alternative.

The key finding is that this framework, called Narrative-to-Action, produces human mobility trajectories that closely match real-world patterns in both spatial and temporal dimensions. It achieves this by integrating large language models (LLMs) in a multi-layer hierarchy: high-level narrative generation, mid-level planning with dynamic adjustments, and low-level execution of actions like location selection and transportation mode choice. This approach moves beyond traditional methods that treat mobility as mere coordinate sequences, instead embedding logical coherence and adaptability into simulated behaviors.

Methodologically, the framework employs a two-stage process at the macro level. First, an LLM acts as a 'creative writer' to generate diary-style narratives based on a character profile, such as 'I woke up at 7:30 AM and grabbed coffee before checking emails.' Then, another LLM parses this narrative into a machine-readable plan with activities, start times, and durations. At the meso level, a reflective decision module uses a novel metric, Mobility Entropy of Occupation (MEO), to introduce probabilistic adjustments—for example, a business owner with high MEO (0.7) might change plans more flexibly than a factory worker with low MEO (0.3). At the micro level, specialized modules ground abstract plans into concrete actions, selecting locations from points of interest using a gravity model and choosing transportation modes based on context.

Results from evaluations on a Guangzhou dataset show that the full framework achieves the highest overall fidelity score of 0.672 in matching real trajectories, outperforming ablated versions. For instance, without the reflective module, mode choice divergence increased from 0.126 to 0.254, indicating poorer alignment with real-world behavior. Qualitative analysis of sample diaries reveals distinct mobility patterns: one depicts a home-centered individual with minimal travel, while another shows a university lecturer with a structured, spatially diverse routine involving commuting and campus activities. These examples demonstrate the system's ability to reproduce heterogeneity in daily schedules, aligning with observed differences in occupation and lifestyle.

This advancement matters for real-world applications in urban governance, such as privacy-preserving data generation for traffic management and infrastructure planning. By simulating how different groups adapt to policies—like congestion pricing—it enables equity-focused interventions without risking personal data. For instance, the framework could predict that professionals with flexible schedules might shift travel times, while workers with rigid routines bear higher costs, informing more inclusive urban designs.

Limitations include the dataset's scale and cultural specificity, which may affect generalizability. Future work should validate the approach with diverse populations and incorporate social interactions, such as family or colleague influences, to enhance realism. Additionally, while the current implementation uses efficient models like GLM-4 Flash for scalability, exploring specialized, smaller models could reduce costs while maintaining performance.