Generalizing Trajectory Intelligence Across Geographic Domains

This direction is merged into another direction: domain shift in trajectory prediction. Its last version stays readable and citable here.

Version 1 of 130 papers36 conceptsfrozen 2026-10-04

When we train a model to predict how a car moves in New York, we aren't just teaching it physics; we are teaching it the specific "grammar" of New York’s streets—its grid layout, its aggressive lane changes, and its unique traffic light timings. This creates a fundamental problem in spatiotemporal AI: models often fail to generalize when moved to a new city. A model optimized for the predictable blocks of Manhattan often founders in the roundabouts of Paris or the sprawling interchanges of Los Angeles.

Recent research has shifted from building site-specific models to developing Generalizable Trajectory Intelligence. The goal is to decouple the fundamental principles of motion from the local "noise" of specific geographic environments.

The Challenge of Cross-City Migration

The primary barrier to universal trajectory models is what researchers call "cross-city migration." As noted by Yichen Liu and colleagues in their work on TrajMamba, trajectory models experience significant performance drops when deployed in cities different from their training data. This is because road network topologies and point-of-interest (POI) distributions are highly localized.

Xie Yu and the team behind BIGCity argue that this necessitates a "foundation model" approach. Instead of training on isolated datasets, they propose unifying individual-level trajectories and population-level traffic states into a single representation. By training on diverse urban data, these models can learn a "universal urban intelligence" that recognizes shared patterns across different geographic regions.

Joint Modeling of Motion and Structure

To make representations more robust, researchers are moving away from treating trajectories as simple sequences of GPS coordinates. Instead, they are grounding motion within the static constraints of the environment.

  • Micro and Macro Fusion: Zhipeng Ma and Zheyan Tu developed JGRM, which treats a path as a multimodal observation. By fusing raw GPS "micro-motion" with map-matched "macro-semantics" (road segments), the model learns representations that are less sensitive to the specific coordinate system of a city.
  • Graph-Based Interactions: Jie Zhao and Chao Chen introduced UniTR, which jointly learns representations for road networks and trajectories. This captures the "many-to-many" interaction between the static road structure and dynamic mobility, helping the model understand why a vehicle moves the way it does based on the available infrastructure.

Learning Universal Rules of Movement

If a model only learns the statistical quirks of a specific dataset, it is essentially memorizing a map. To achieve "plug-and-play" capability, models must learn the underlying causal mechanisms of movement.

Xiaohe Li and Feilong Huang address this through MetaTra, a meta-learning framework. By simulating domain shifts during training, the model learns to adapt to unseen scenarios—like moving from a quiet library to a bustling street—without needing new data. Similarly, Shiji Song uses causal intervention in SEAD to adjust for "social environment" confounders, ensuring the model doesn't rely on scene-specific biases.

In the domain of autonomous driving, Daniel Grimm and Ahmed Abouelazm propose HoliGraph:Goal, which "softly binds" predicted paths to the road network. By focusing on goal-based navigation rather than just coordinate extrapolation, the system maintains high accuracy even when the underlying road geometry is unfamiliar.

Bridging the Data Gap

When real-world data is missing for a new environment, researchers are turning to synthetic generation. John Palowitch and Jason Baldridge demonstrated in Into the Unknown that using Large Language Models (LLMs) to generate synthetic, map-grounded instructions can improve navigation in unseen cities by over 45%. This suggests that the future of generalizable trajectory intelligence lies in combining massive, diverse datasets—like the Amelia airport dataset or the L4 Motion Forecasting benchmark—with architectures that can bridge the gap between general linguistic/spatial understanding and local geographic features.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generalizing_trajectory_intelligence_acr_5990a38d,
  title = {Generalizing Trajectory Intelligence Across Geographic Domains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalizing-trajectory-intelligence-across-geographic-domains/rdr_5990a38db6794f36b9585b3cf3086483},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_generalizing_trajectory_intelligence_acr_5990a38d_v1,
  title = {Generalizing Trajectory Intelligence Across Geographic Domains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalizing-trajectory-intelligence-across-geographic-domains/rdr_5990a38db6794f36b9585b3cf3086483/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}