Deep learning models excel at predicting vehicle trajectories, forecasting traffic flows, and optimizing fleet routing when they are evaluated in the exact same environment where they were trained. However, drop that same model into a new city, and its performance often collapses. A model trained on the rigid grid system of Manhattan will struggle to make sense of the radial, curvilinear streets of Paris.
As highlighted by Ye et al. and Jacobs et al., this lack of spatial transferability creates a massive bottleneck, forcing practitioners to undergo costly data collection and localized retraining for every new geographic deployment. To build truly universal urban intelligence, researchers are shifting away from localized architectures toward city-agnostic frameworks that generalize across geographic domain shifts.
The Root Causes of Geographic Brittleness
To understand why spatial models fail to generalize, we must look at what they actually memorize during training. The brittleness stems from two primary types of overfitting:
1. Coordinate and Topological Overfitting
Many trajectory models rely on absolute spatial coordinates (latitude and longitude) or location offsets relative to a specific central reference point. As Buchholz points out, this ties the model's internal representations directly to a single geographic region. Furthermore, models frequently memorize specific road network topologies and localized driving cultures rather than universal motion principles, a vulnerability documented by Liu et al. and Tong et al.. When the underlying spatial constraints change, the model encounters a severe distribution shift.
2. The Vocabulary Mismatch Problem
In many spatial architectures, road networks are discretized into unique, localized identifiers (such as specific road segment IDs or grid cells). As explained in Grid and Road Expressions Are Complementary for Trajectory Representation Learning, this creates a fundamental vocabulary mismatch and a "cold start" problem. Because road segment tokens are discrete and unique to each city, a model cannot easily map its learned navigation logic to the completely unobserved infrastructure tokens of a new city.
Traditional Approach (Brittle):
[Absolute Coordinates / City-Specific Road IDs] ──> Overfits to Local Topology ──> Fails in New City
Generalizable Approach:
[Shape-Agnostic / Universal Semantic Spaces] ──> Extracts Invariant Motion Dynamics ──> Transfers Zero-Shot
Strategy 1: Coordinate-Agnostic and Shape-Based Representations
The first step in building cross-city generalization is stripping away absolute spatial coordinates. If a model does not know its absolute position on a global map, it is forced to learn the geometric and structural relationships of movement.
Landi addresses this by moving away from coordinate-dependent features toward "Geolets"—discriminative sub-sequences tailored to capture scale-invariant dynamics of movement. To make this geometric abstraction actionable, Landi et al. introduced a shape-based map matching approach that completely ignores absolute latitude and longitude. By representing road networks as a symbolic "alphabet" of turns and curves, the model identifies identical geometric behaviors across entirely different cities. A right-angle turn followed by a sharp deceleration means the same thing semantically, whether it occurs in Rome or Tokyo.
Strategy 2: Bridging the Vocabulary Gap with Universal Embeddings
To solve the vocabulary mismatch caused by city-specific road identifiers, researchers are replacing discrete tokens with continuous, universally accessible semantic spaces.
Instead of assigning an arbitrary ID to a street, Wei et al. developed a framework called TransferTraj that encodes spatial context through universal text embeddings. By translating discrete infrastructure into descriptive language, the model can reason about a "four-lane highway merge" or a "residential cul-de-sac" across different cities without requiring any architectural re-initialization.
Taking this linguistic grounding a step further, Liu et al. and Lepri have demonstrated that Large Language Models (LLMs) can be repurposed as zero-shot next-location predictors. By leveraging the internal geographical and spatial reasoning capabilities of LLMs, these frameworks bypass discrete city-specific identifiers entirely, predicting continuous coordinates based on semantic descriptions of human mobility patterns.
Strategy 3: Disentangling Causal Dynamics from Local Quirks
A generalizable model must separate the universal physics of motion from the "spurious" environmental quirks of its training data. For instance, a vehicle slowing down near an intersection is a causal relationship governed by traffic rules; a vehicle slowing down because of a highly specific local billboard is a spurious correlation.
To enforce this distinction, Pourkeshavarz proposed CaDeT, a causal disentanglement framework that uses latent space interventions to isolate invariant causal factors (like road geometry) from coincidental neighbor behaviors. By ignoring misleading local correlations, the model retains its predictive accuracy in unseen environments.
Spatial Features ───┬───> Invariant Causal Factors (Geometry, Physics) ───> Generalized Prediction
└───> Spurious Local Quirks (City-Specific Layouts) ───> Discarded via Intervention
When remaining domain shifts cannot be fully eliminated offline, models must adapt dynamically during inference. Loh et al. utilize domain adversarial training to minimize the statistical discrepancy between different traffic environments. Alternatively, Fujii et al. introduced an In-Context Learning framework for trajectory prediction. Rather than updating model weights via backpropagation, their architecture adapts to the unique geometry and walking behaviors of a novel scene at inference time simply by conditioning its predictions on a few local demonstration examples.
Key Foundational Work
How to Build a Graph-Based Deep Learning Architecture in Traffic Domain: A Survey (2020) — contributed 2 concept elements to this direction.
Supply-Demand Prediction of DiDi Based on Points of Interests Selection in Extreme Gradient Boosting Algorithm (2020) — contributed 1 concept elements to this direction.
PRANK: motion Prediction based on RANKing (2020) — contributed 1 concept elements to this direction.