A shared AI model trained on data from five signalized intersections produced lower 10-second vehicle trajectory errors than models trained separately at each site, according to an arXiv preprint. The advantage appeared at all five sites: median site-level reductions were 36.8% for minADE and 22.0% for minFDE. Summarized through ratio-of-macro-average reductions, the gains were 37.1% and 28.4%.
A common frame for different junctions
That result addresses a practical question: can heterogeneous intersection data support one reusable model without erasing the local context that governs behavior? The approach uses a domain-specific, non-learned harmonization operator. It converts each scene into a common actor-centric context while retaining local geometry and normalized signal state, and it does not require a learned site identifier.
Behind that representation is a two-stage spatial and temporal autoregressive backbone paired with a 25-mode Gaussian mixture model head. The head is trained with masked negative log-likelihood, or NLL, a measure of how well the model's probability estimates fit observed trajectories.
The evaluation uses roadside trajectory and signal data from five signalized intersections in two Florida regions: four sites in Gainesville and site 22 in South Florida. Balanced sampling gave each site 100,000 training windows and equal-sized test sets.
Forecasts were conditional and focused on one vehicle at a time. The model received two seconds of history sampled at 10 Hz, then generated 2-, 5- or 10-second futures, with runs extending to as many as 100 generated steps. Neighboring vehicles, pedestrians and signals were replayed from logs, so the setup was not an interactive traffic simulation.
To judge the forecasts, the researchers used minADE and minFDE, lower-is-better displacement measures, over six sampled rollouts for each scenario. They evaluated the 2-, 5- and 10-second horizons separately and also reported test NLL.
The advantage grew with time
The pooled model's performance was mixed at short range but more consistent over longer rollouts. At two seconds, it improved both measures at four sites and trailed local training only at site 9 by about 2%. At five seconds, it improved both measures at sites 7, 8 and 22, while sites 9 and 10 kept small margins. At 10 seconds, pooled training was ahead at every site.
Transfer helped, but unevenly
Whether the shared model could travel to an unseen site was less straightforward. In the paper's zero-shot test, a model trained without target behavior-model parameter optimization was applied to a held-out intersection. Zero-shot does not mean no target data: deployment still uses the target's movement geometry, stop lines and observed signal context.
At 10 seconds, zero-shot beat local training on both metrics at four of five sites. Site 9 worsened, with reported changes of -14.5% and -21.5%, while the single cross-region fold at South Florida site 22 improved the two errors by 38.2% and 27.7%. At two seconds, however, zero-shot trailed local training everywhere except site 8, with median changes of -36.0% and -41.9%. The comparison does not identify the mechanism behind that horizon dependence.
Fine-tuning offered a selective middle ground. The zero-shot checkpoint was updated with 1,000 target windows, equal to 1% of the 100,000-window local budget. At 10 seconds, this improved on zero-shot at sites 8, 9 and 22, was essentially unchanged at site 7 and worsened site 10. Site 8 was the best regime across every horizon and metric, and site 22 recorded the best 10-second minADE. Because there was no matched 1,000-window scratch control, the isolated value of pretrained initialization is not identified.
A fixed-volume test adds evidence
A fixed-volume experiment at site 7 pointed in the same direction. With the budget held at 100,000 windows, every tested cross-site mixture improved every horizon over pure-local training. A mixture that replaced 20% of local windows with equal shares from other sites had 66.2% lower minADE and 58.4% lower minFDE at 10 seconds. Test NLL also improved: two tested mixtures recorded -5.97 and -5.82, compared with -4.02 for pure local.
The comparison with a classical baseline added a caution. After long autoregressive rollouts, local models fell behind calibrated IDM at the two highest-flow sites, whereas pretrained-backbone regimes did not in the reported runs. The displacement score takes the best of six sampled trajectories, which can favor probabilistic models relative to deterministic baselines. The improved test NLL in the site 7 sweep provides a second signal, but it does not remove that scoring caveat.
The result has a clear boundary
The boundaries of the evidence are important. The test covers five Florida signalized intersections, and the cross-region result rests on one South Florida fold. The forecasts are replay-conditioned and single-vehicle: surrounding actors and signals do not respond interactively to the generated vehicle. The study therefore evaluates conditional trajectory generation in this setup, not an interactive traffic simulation.
Taken together, the reported pooled results were strongest and most consistent at the longest tested horizon, while zero-shot reuse was less reliable at two seconds and fine-tuning varied by site. The evidence concerns model reuse for replay-conditioned single-vehicle generation across these five intersections, rather than a general claim about traffic behavior.
Paper data and sources
Original title: PhaseShift: Topology-Aware Data Harmonization and Model Consolidation Across Signalized Intersections
Authors: Yash Ranjan, Artur Kumik, Rahul Sengupta et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text