💻 computer science

PhaseShift: Topology-Aware Data Harmonization and Model Consolidation Across Signalized Intersections

PhaseShift is a topology-aware framework that harmonizes heterogeneous traffic data across signalized intersections into a shared actor-centric representation, enabling a single reusable backbone model to outperform site-specific local models in trajectory prediction accuracy and generalization while supporting zero-shot deployment and low-data adaptation.

Original authors: Yash Ranjan, Artur Kumik, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka

Published 2026-08-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yash Ranjan, Artur Kumik, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, millions of vehicles navigate a complex web of roads, guided by traffic lights, road markings, and the unpredictable actions of other drivers. To make these journeys safer and more efficient, engineers often rely on computer models that learn how cars behave. Traditionally, these models are built like custom suits: a unique model is trained for every single intersection, learning the specific patterns of that one corner. This approach works, but it is incredibly expensive and inefficient. If a city has hundreds of intersections, each with different traffic lights, road layouts, and driver habits, engineers must build and maintain hundreds of separate models. Furthermore, a model trained at one busy downtown corner cannot easily help a quiet suburban intersection, even if the basic rules of driving are the same. The challenge lies in creating a single, smart system that can understand the universal rules of traffic while still respecting the unique details of each location.

Researchers at the University of Florida have tackled this problem with a new approach called PhaseShift. Instead of training a separate model for every intersection, they developed a method to teach one central model how to drive at five different intersections in Florida. The team collected real-world data from five signalized intersections, capturing the movements of vehicles, pedestrians, and the changing colors of traffic lights. These locations varied significantly: some were wide, busy arterial roads with dozens of turning lanes, while others were smaller, quieter crossings. Some had traffic lights that changed quickly, while others had long cycles. The goal was to see if a single computer brain could learn from all these different environments and then apply that knowledge to predict how a car would move, even at an intersection it had never seen before.

To make this possible, the researchers first had to solve a language barrier. Data collected at one intersection looks very different from data at another. One camera might measure distance from a specific pole, while another uses a different starting point. One traffic light controller might use specific codes for its phases, while another uses a completely different system. If a computer tries to learn from these raw, messy inputs, it gets confused by the differences in measurement rather than learning the actual behavior of the drivers. The team created a translation system that converts all these different observations into a common language. They stripped away the local quirks, such as where the camera was placed or what the traffic light was called, and focused on what actually matters: where the car is relative to the driver, where the stop lines are, what the traffic light is doing, and how other vehicles are moving. This process allowed them to feed data from five very different places into a single training program.

The result was a shared model that learned the general rules of driving across all five sites. When tested, this single model performed remarkably well. At a ten-second horizon, which is a significant amount of time in traffic terms, the shared model predicted vehicle movements more accurately than the five separate, custom-trained models at every single location. On average, it reduced prediction errors by more than a third compared to the local models. This suggests that the model had successfully learned the underlying physics of traffic flow and could apply those lessons universally. Even more impressively, when the researchers tested the model on a new intersection it had never seen during training, it still performed better than a model trained from scratch on just that one new location. This "zero-shot" capability means the system can be deployed to a new city or a new intersection with very little additional work, simply by feeding it the local road layout and traffic light patterns.

However, the study also revealed that a one-size-fits-all approach is not always perfect. While the shared model excelled at predicting long-term movements, it sometimes struggled with very short-term predictions at specific sites, particularly those with unusual traffic patterns or signal timings. For these tricky locations, the researchers found that a small amount of fine-tuning—retraining the shared model with just a tiny fraction of local data—could bridge the gap. This hybrid approach, where a strong general model is slightly adjusted for local quirks, proved to be the most effective strategy. It allowed the system to benefit from the vast amount of data collected across all sites while still adapting to the unique needs of a specific corner.

The researchers also discovered something surprising about how these models behave over time. When predicting the path of a car for a long duration, the local models trained on a single site began to fail, making errors that grew larger and larger until the predictions were worse than a simple, rule-based calculation. The shared model, however, did not suffer from this collapse. It remained stable and accurate, suggesting that learning from a variety of traffic conditions helps the system understand the fundamental rules of driving better than focusing on just one location. This stability is crucial for safety, as it means the model is less likely to make catastrophic mistakes when predicting where a car will be in the future.

Ultimately, this work demonstrates that traffic data from different places can be combined to create smarter, more robust models. By harmonizing the data and focusing on the shared physical reality of driving, the researchers showed that we do not need a unique model for every intersection. Instead, we can build a single, adaptable system that learns from the collective experience of many roads. This approach not only saves time and resources but also produces models that are more reliable and better at handling the unexpected. As cities continue to grow and traffic becomes more complex, the ability to share knowledge across different locations will be essential for building safer and more efficient transportation networks. The study confirms that while every intersection has its own personality, the language of traffic is universal, and a single model can learn to speak it fluently.

Technical Summary: PhaseShift

Problem Statement
Deploying learned traffic-behavior models across a road network often results in a portfolio of narrowly specialized models, where each intersection and operating condition is treated as an independent dataset. This approach is costly to train, calibrate, and maintain, and it prevents evidence collected at one site from strengthening models at another. The challenge is particularly acute at signalized intersections, where observed trajectories arise from an entangled physical system involving topology, signal control, demand, and heterogeneous human responses. Furthermore, roadside sensing introduces distinct coordinate frames, controller conventions, and tracking noise. Existing methods either rely on manually authored road networks (classical simulators) or are developed around specific geometries, failing to address the dataset-harmonization problem required for cross-site learning from noisy, real-world measurements.

Methodology
The paper introduces PhaseShift, a topology-aware framework designed to harmonize heterogeneous roadside trajectories into a shared, actor-centric representation and train a single reusable backbone.

Data Harmonization: PhaseShift converts raw roadside scenes into a common token space without using learned site identifiers. It employs a non-learned harmonization operator that:
- Uses ego-relative coordinates to ensure global-pose invariance (removing site orientation and origin).
- Utilizes trajectory-induced movement paths instead of authored lane graphs. These paths are derived from collected tracks, normalized by arc length, and summarized into representative polylines.
- Incorporates normalized signal context, converting controller-specific phase identifiers into site-independent categorical states (e.g., red, protected green) attached to geometric elements.
- Represents entities (actors, pedestrians, stop lines) as variable-cardinality tokens with type-specific descriptors, accommodating incomplete observations common in roadside tracking.
Shared Backbone Architecture: The model consists of two stages:
1. Spatial Encoding: A per-timestep encoder aggregates variable-cardinality scene tokens (neighbors, lane geometry, signals) into a single ego embedding using additive composition and neighbor cross-attention.
2. Temporal Decoding: An autoregressive decoder predicts a multimodal next-state distribution (using a Gaussian Mixture Model with 25 modes). It employs relative position encoding (RPE) to inject spatial displacement and elapsed time without exposing a global coordinate frame.
Training and Deployment Lifecycle: The shared parameters ( $\theta^\star$ $θ^{⋆}$ ) are pretrained across multiple domains using a masked negative log-likelihood objective. The framework supports three deployment regimes:
- Pooled: One shared model over all known domains.
- Zero-Shot: Applying the pretrained backbone to a held-out intersection without target-site parameter optimization.
- Fine-Tuned: Initializing from the zero-shot checkpoint and updating with a small set of target windows (1,000 samples).

Key Contributions

Topology- and Control-Aware Data Harmonization: The framework maps diverse trajectories, movement geometries, and signal states into a shared token space while preserving the local physical-control structure necessary for conditional generation.
Shared Pretraining and Specialization: A single backbone supports pooled operation, zero-shot deployment, and low-data adaptation. The framework uses per-domain residuals to distinguish sites covered by the common model from those requiring specialization.
Five-Domain Field Evaluation: The authors evaluate the approach on balanced field data from five intersections in two Florida regions (Gainesville and a South Florida arterial), comparing pooled, zero-shot, and fine-tuned regimes against independently trained local models and classical baselines (IDM, constant acceleration).

Results
The evaluation, conducted under a replay-conditioned, best-of-sampled-trajectory protocol, yields the following findings:

Pooled Performance: At a 10-second horizon, the pooled model lowers both minADE and minFDE relative to independently trained local models at all five sites. Median reductions are 36.8% for minADE and 22.0% for minFDE.
Zero-Shot Deployment: Leave-one-intersection-out deployment (including a cross-region fold) outperforms local training on both 10-second metrics at four of the five sites. However, short-horizon (2-second) performance is less uniform, with zero-shot models often trailing local models.
Fine-Tuning: Fine-tuning with 1,000 target update windows improves upon zero-shot performance at three sites and is the strongest regime at one site (Site 8).
Long-Horizon Stability: At the two highest-flow sites (7 and 22), locally trained models fall behind a calibrated Intelligent Driver Model (IDM) after long autoregressive rollouts. Pretrained-backbone regimes (Pooled, Zero-Shot, Fine-Tuned) avoid this reversal, maintaining superior performance.
Composition vs. Volume: A fixed-window composition sweep at Site 7 demonstrates that mixing data from other sites (even with a fixed total window count of 100,000) sharply lowers long-horizon error compared to pure local training. Test-likelihood gains argue against a "best-of-N dispersion-only" explanation.

Significance and Claims
The paper claims that PhaseShift demonstrates the feasibility of consolidating heterogeneous physical-control settings into a stronger shared long-horizon model while retaining the ability to identify sites that require adaptation. The authors emphasize that the protocol measures conditional single-vehicle generation under replayed context, not closed-loop traffic simulation.

The significance lies in shifting from a "train-from-scratch" default to a lifecycle of consolidation: harmonizing data to learn a reusable backbone, deploying it directly where it generalizes, and adapting only domains with meaningful residuals. The results suggest that for regional deployment, a single pooled model can outperform a portfolio of specialized models, particularly in long-horizon prediction and cross-region generalization, provided that the harmonization preserves the necessary physical and control context. The authors remain modest, noting that the study does not isolate the specific causal factors (topology vs. demand vs. control) driving the improvements and that further stratification and ablation studies are needed to fully attribute the gains.

Technical Summary: PhaseShift

More like this