PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yash Ranjan, Artur Kumik, Rahul Sengupta, Anand Rangarajan, Sanjay Ranka

arXiv:2608.25275v1cs.AIcs.RO

TL;DR

Site-specific traffic models limit evidence reuse across heterogeneous intersections. PhaseShift harmonizes roadside trajectories into a shared actor-centric representation and trains a reusable backbone for pooled, zero-shot, and low-data adapted deployment. In the five-site evaluation, pooled training improved both 10-s displacement metrics over local training at every site, while the protocol remained conditional generation under replayed context rather than closed-loop simulation.

  • Problem

    Traffic-behavior models are commonly trained separately for each intersection, limiting principled reuse across heterogeneous physical, control, demand, and sensing regimes.

    Source: §I · p. 1

  • Method

    PhaseShift harmonizes trajectories, geometry, signal state, and actor context into a shared actor-centric space, then pretrains one actor-conditional backbone for pooled use, zero-shot transfer, or fine-tuning.

    Source: §I · p. 1, A. Stage 1 — Spatial encoding · p. 4

  • Results

    At 10 s, POOLED improves both displacement metrics over LOCAL at all five sites, with median site-level reductions of 36.8% for minADE and 22.0% for minFDE.

    Source: C. One shared model, held-out… · p. 6

  • Takeaways & Limitations

    Within this five-site evaluation, one shared checkpoint supports consolidation across heterogeneous intersections, while residual performance identifies sites where adaptation remains useful.

    Source: §IX · p. 9, §XI · p. 9

    The evaluation uses one checkpoint per learned configuration, limited temporal evidence, and a replay-conditioned best-of-six protocol rather than closed-loop simulation.

    Source: §X · p. 9

Abstract

from arXiv · show

Learned traffic-behavior models are commonly trained separately for each intersection, creating model portfolios that cannot share evidence across sites. We present PhaseShift, a topology-aware framework that harmonizes heterogeneous roadside trajectories into a shared actor-centric representation and trains one reusable backbone. Ego-relative coordinates, trajectory-induced movement paths, normalized signal context, and variable-cardinality interaction tokens remove site conventions while preserving behaviorally relevant topology. The backbone supports pooled operation, zero-shot at a held-out intersection, and low-data adaptation. We evaluate five intersections in two Florida regions on balanced field data, 100k training windows and equal-sized test sets per site under a replay-conditioned, best-of-sampled-trajectory protocol. At 10s, one pooled model lowers both minADE and minFDE relative to trained local models at all five sites, with median reductions of 36.8% and 22.0%. Leave-one-intersection-out deployment, including one cross-region fold, beats local training on both 10-s metrics at four of five sites, although short-horizon performance is less uniform. Fine-tuning with 1,000 target update windows improves on zero-shot at three sites and is the strongest regime at one. At site 7, every cross-site mixture sharply lowers long-horizon error under a fixed 100k-window budget; test-likelihood gains argue against a best-of-sample dispersion-only explanation. Local models fall behind calibrated IDM at the two highest-flow sites after long autoregressive rollouts; pretrained-backbone regimes do not. Within this five-site evaluation, PhaseShift demonstrates consolidation across heterogeneous physical control settings while identifying sites that still require adaptation. The protocol measures conditional single-vehicle generation under replayed context, not closed-loop traffic simulation.

I. INTRODUCTION

PhaseShift addresses the cost and limited reuse of site-specific traffic models by harmonizing heterogeneous intersection observations into a shared actor-centric representation. It trains one reusable backbone for pooled operation, held-out deployment, and selective adaptation while preserving local physical and control context.

Motivation

  • Up to 35 models can be required when five intersections across seven day-of-week domains are treated independently.Separate training, calibration, validation, and maintenance prevent observations from one domain from strengthening another.Source: §I · p. 1
  • Cross-site learning must remove coordinate and controller conventions while retaining topology, signal control, demand, interactions, and sensing context.These factors jointly generate observed behavior at signalized intersections.Source: §I · p. 1

Contributions

  • PhaseShift maps roadside trajectories, neighboring actors, pedestrians, movement geometry, stop lines, and signal states into a shared actor-centric token space.The representation is designed for conditional trajectory generation without erasing local physical-control structure.Source: §I · pp. 1–2
  • One shared backbone supports pooled operation, leave-one-intersection-out zero-shot deployment, and low-data adaptation without a learned site identifier.Residuals can indicate where specialization or investigation is warranted.Source: §I · pp. 1–2

Evaluation

  • The evaluation covers balanced field trajectories from five intersections in two Florida regions, comparing pooled, local, held-out, adapted, and fixed-window mixture regimes.The composition sweep controls nominal window count at one site as an initial check on the diversity–volume confound.Source: §I · p. 2

III. MULTI-INTERSECTION PRETRAINING FORMULATION

PhaseShift defines domains as joint physical-control regimes and applies a non-learned harmonization operator before shared actor-conditional pretraining. The resulting parameters support pooled use, zero-shot transfer with target context, or fine-tuning for residual target behavior.

Domain formulation

  • A domain may represent an intersection or an intersection under a recurring operating condition, with each domain treated as a joint physical-control regime.Raw scenes differ in coordinates, movement paths, signal identifiers, and visible actor sets.Source: §III · p. 2
  • The non-learned harmonization operator hm converts each scene into a common actor-centric context using local geometry and normalized signal state.This establishes a shared input space before behavioral transfer.Source: §III · p. 2

Pretraining

  • Shared parameters are pretrained over harmonized domains, with wm controlling domain sampling and fθ denoting the shared actor-conditional model.No learned site identifier is required.Source: §III · p. 2

Deployment

  • For an unseen target, its geometry and signal context are supplied to the pretrained model for zero-shot use, or the model is initialized from θ⋆ and fine-tuned with target data.Zero-shot transfer therefore uses target context even without target behavior-model parameter optimization.Source: §III · p. 2

Conditional generation

  • The transition model conditions on a target actor’s harmonized history, valid neighboring observations, movement geometry, stop lines, and signal state.Demand is represented through evolving actor and movement occupancy rather than a learned site label.Source: §III · p. 2
  • Generated target states are fed back autoregressively while neighboring actors, pedestrians, and signal state are replayed from the logged scene.The formulation separates harmonization, shared behavior learning, and optional residual adaptation.Source: §III · p. 2

IV. CROSS-DOMAIN DATA HARMONIZATION

Cross-domain harmonization expresses roadside observations in a shared token space while preserving geometry, control state, and actor context. It uses trajectory-induced movement geometry and variable-cardinality observations without claiming robustness beyond the reported field evaluation.

Representation

  • The harmonizer removes global coordinate and orientation conventions while preserving local geometry and structured behavioral variation.It also avoids fixed actor ordering and learned site identifiers.Source: §IV · p. 3
  • Ego-relative tokens encode each entity’s pose and type-specific dynamics or extent, making globally translated or rotated scenes equivalent.This invariance does not extend to local geometry itself.Source: §IV · p. 3
  • Token descriptors separately encode neighbor actors, lane geometry, stop lines, and normalized signal state.Signal state is attached to the governed geometric element rather than represented by controller-specific identifiers.Source: §IV · p. 3

Movement geometry

  • Trajectory-induced movement polylines replace authored lane-connectivity graphs and are built from sufficiently complete tracks using movement grouping and arc-length normalization.Representative paths are aligned and discretized, while under-supported movement classes are filtered out.Source: §IV · p. 3

Data separation

  • The geometry-building interval is disjoint from behavioral training, validation, and test data, while behavioral inference can consume arbitrary partial tracks.The study does not quantify the per-site deployment cost of constructing geometry.Source: §IV · p. 3

Observation handling

  • Variable-cardinality encoding removes tokens for currently invalid observations instead of imputing a dense fixed-size actor tensor.This accommodates changing roadside tracks without asserting robustness beyond the field evaluation.Source: §IV · p. 3

V. SHARED BACKBONE AND DEPLOYMENT LIFECYCLE

PhaseShift combines per-timestep spatial encoding with autoregressive temporal decoding inside a reusable multi-intersection lifecycle. The shared backbone is trained once across harmonized domains and then pooled, deployed zero-shot, or fine-tuned as needed.

Backbone

  • PhaseShift encodes each harmonized scene spatially at every timestep, then decodes a multimodal next-state distribution autoregressively.The spatial encoder aggregates a variable-cardinality scene into an ego embedding.Source: §V · p. 3

Stage 1 — Spatial encoding

  • Stage 1 uses weights shared across the sequence to produce timestep-level scene representations.These representations feed the temporal decoder.Source: A. Stage 1 — Spatial encoding · p. 3
  • Each modality is projected into a common dmodel = 256 space, with shared actor-state and map-speed tokenizers and type-specific relative-pose encoders.The tokenization blocks use Linear, LeakyReLU, and Dropout components.Source: A. Stage 1 — Spatial encoding · p. 3
  • Additive token composition combines modality encodings and adds a signal term for signal tokens.The factorization separates agent behavior from relative position.Source: A. Stage 1 — Spatial encoding · p. 3

Deployment lifecycle

  • The lifecycle reuses the shared backbone as a pooled model, deploys it zero-shot at a new intersection, or fine-tunes it when residual error indicates specialization may be needed.Residual error is a prioritization signal, not proof of a specific behavioral cause.Source: Figure 1 · p. 4

Stage 1 — Spatial encoding

  • Neighbor attention queries with the ego embedding and attends over neighbor, lane-polyline, and signal tokens to form one scene-aware ego embedding per timestep.The output has shape [batch, seq_len, dmodel].Source: A. Stage 1 — Spatial encoding · p. 4

B. Stage 2 — Temporal decoding

Stage 2 decodes future motion autoregressively by attending to scene-aware history and producing multimodal displacement distributions.

VI. DATA AND MULTI-INTERSECTION PROTOCOL

PhaseShift evaluates five heterogeneous signalized intersections using balanced, chronologically separated field data and site-level or held-out-domain protocols.

  • Five intersections span four Gainesville sites and one high-volume South Florida arterial intersection used as the cross-region target.Source: §VI · p. 4
  • Each intersection is treated as a pretraining domain, with roadside tracks generated from 10-Hz fisheye video and mapped into a rectangular ground frame.Source: §VI · p. 4
  • The domains vary in recorded duration, demand, movement count, and signal operation, while Gainesville sites share regional driving norms absent at site 22.Source: §VI · p. 4
  • The aerial layouts differ in footprint, approach width, lane arrangement, turn movements, and crosswalk placement, so the views are illustrative rather than metrically aligned.Source: §VI · pp. 4–5, Figure 2 · p. 5
  • Each site contributes 100,000 training samples, and chronological partitioning prevents tracks or overlapping windows from crossing evaluation boundaries.Source: §VI · p. 5
  • POOLED trains on all five domains, whereas ZERO-SHOT leaves one intersection out and applies the resulting model without target-site parameter optimization.Source: §VI · p. 5
  • Checkpoint selection uses validation NLL without the target test interval, although zero-shot checkpoint selection still uses target validation data.Source: §VI · p. 5

VII. PRETRAINING AND DEPLOYMENT REGIMES

PhaseShift compares local, pooled, zero-shot, and fine-tuned shared architectures under conditional single-vehicle rollout evaluation, emphasizing long-horizon consolidation and held-out deployment.

  • PhaseShift seeks one common backbone plus specialization only for domains whose validation results support it, rather than one parameter set per intersection.Source: §VII · p. 5
  • LOCAL trains from scratch, POOLED trains across known domains, ZERO-SHOT excludes the target during pretraining, and FINE-TUNED updates zero-shot initialization with 1,000 target windows.Source: §VII · pp. 5–6
  • The protocol evaluates conditional single-vehicle rollouts with replayed neighbors, pedestrians, and signal state rather than fully interactive traffic simulation.Source: A. Conditional trajectory-rollout… · p. 6
  • Figure 3 reports shared-backbone error reductions relative to LOCAL, where positive values indicate lower shared-model error and negative values indicate a specialization gap.Source: Figure 3 · p. 6, C. One shared model, held-out… · p. 6
  • At 10 s, POOLED lowers minADE and minFDE relative to LOCAL at all five sites, with median reductions of 36.8% and 22.0%.Source: C. One shared model, held-out… · p. 6
  • At 10 s, ZERO-SHOT beats LOCAL on both metrics at four of five sites, while one cross-region fold reduces errors by 38.2%/27.7%.Source: C. One shared model, held-out… · p. 6
  • At 2 s, ZERO-SHOT trails LOCAL everywhere except site 8, with median changes of −36.0%/−41.9%.Source: C. One shared model, held-out… · p. 7

D. Target fine-tuning across intersections

Fine-tuning uses 1,000 target windows to adapt zero-shot checkpoints, improving long-horizon performance at several sites but remaining horizon- and site-dependent. A fixed-volume site-7 sweep shows that cross-site composition, rather than simply more windows, drives large gains, although temporal coverage and sampling effects remain unresolved.

Target fine-tuning

  • 1,000 target windows improve both zero-shot metrics at sites 8, 9, and 22, leave site 7 essentially unchanged, and degrade site 10 at 10 s.Site 10 minADE changes from 0.8805 to 1.1256.Source: D. Target fine-tuning across… · p. 7
  • 38.2%/30.8% are the median 10-s minADE/minFDE reductions of FINE-TUNED relative to LOCAL across the five sites.The site-10 minFDE comparison is the exception, with a −11.1% change.Source: D. Target fine-tuning across… · p. 7

Fixed-volume composition sweep

  • At fixed site-7 training volume, replacing 20% of local windows with equal shares from the other four sites cuts 10-s error by 66.2%/58.4%.The sweep uses 100,000 total windows, with 100/0/0/0/0 representing LOCAL.Source: E. Fixed-volume composition sweep · pp. 7–8, Table III · p. 8
  • The 80/5/5/5/5 mixture is best at 10 s and comes within 2.8%/4.0% of POOLED despite using one fifth as many total windows.The decisive contrast is between pure-local and mixed-site training; more non-local data is not uniformly better.Source: E. Fixed-volume composition sweep · p. 8
  • The composition sweep fixes window count but not independent temporal evidence, and best-of-N scoring may reward wider output dispersion.At 10 s, improved test NLL argues against a dispersion-only explanation, while additional controls are still needed before generalization.Source: E. Fixed-volume composition sweep · p. 8

F. Classical baselines and long-horizon behavior

Learned models outperform classical baselines at short horizons, but local models can lose their advantage during long autoregressive rollouts at the two highest-flow sites. Cross-domain pretrained regimes avoid that observed reversal, while likelihood results support genuine long-horizon gains and site-specific specialization remains evident.

Classical baselines

  • At 2 and 5 s, every learned regime has substantially lower best-of-sample error than CONST.-VEL. and IDM at every site.The comparison is subject to the unequal hypothesis budget noted in the evaluation.Source: F. Classical baselines and… · p. 8

Long-horizon behavior

  • At 10 s, LOCAL falls behind both classical baselines at site 7 and behind IDM at site 22, the two highest-flow sites.At site 7, LOCAL records 2.1595/5.1121 versus IDM’s 1.7216/4.0604.Source: F. Classical baselines and… · p. 8

Likelihood

  • At 10 s, a pretrained-backbone regime has the best NLL at every site, including −5.94 versus −4.02 for LOCAL at site 7.The result argues against long-horizon displacement gains being solely an artifact of wider output dispersion under best-of-N scoring.Source: F. Classical baselines and… · p. 8

Per-site synthesis

  • 67.1%/60.0% are site 7’s POOLED reductions in 10-s minADE/minFDE, while every cross-domain regime has much lower 10-s error than LOCAL.Site 7 is the busiest Gainesville domain and shows the sharpest long-horizon failure.Source: G. Per-site synthesis · p. 8, F. Classical baselines and… · p. 8
  • Site 22’s local 10-s error of 1.80/4.56 improves to 1.11/3.29 with ZERO-SHOT and 0.96/2.69 with POOLED in the cross-region case.FINE-TUNED and POOLED split the best 10-s metrics, and both decisively outperform local training.Source: G. Per-site synthesis · p. 8
  • Site 9 is the clearest specialization signal: ZERO-SHOT degrades both 10-s metrics, POOLED concedes short-horizon accuracy, yet POOLED remains best at 10 s.The site has atypical signal timing and the lowest flow among the five intersections.Source: G. Per-site synthesis · p. 8
  • At site 10, POOLED wins at 2 and 10 s, whereas 10-s fine-tuning increases minADE from 0.88 to 1.13 versus ZERO-SHOT.Across sites, the backbone supports consolidation while site 9 still signals horizon-specific specialization.Source: G. Per-site synthesis · pp. 8–9

IX. DISCUSSION

PhaseShift supports consolidation across five heterogeneous intersections while retaining selective adaptation, but the evidence remains bounded by the replay-conditioned evaluation and unresolved attribution and selection questions.

Results

  • At site 7, every fixed-volume cross-site mixture sharply outperforms pure-local training, indicating that mixture composition matters beyond nominal window count.The sweep fixes window count but not independent temporal evidence, and wider dispersion remains a possible influence under best-of-N scoring.Source: §IX · p. 9, §X · p. 9

Operational consolidation

  • At 10 s, POOLED is the best learned regime at sites 7, 9, and 10, while FINE-TUNED wins both metrics at site 8.POOLED is within 5% of the per-metric best at site 22, but this assignment was chosen from test results rather than a validated deployment policy.Source: §IX · p. 9

Scope and limitations

  • The evaluation does not execute the proposed validation-driven policy for consolidating known domains and specializing only when residuals warrant it.Operational selection must use validation under one fixed objective rather than test results.Source: §IX · p. 9

Physical-AI perspective

  • PhaseShift frames harmonization as an interface between learned models and infrastructure because layout, control, demand, behavior, and sensing jointly shape observations.The study bridges five realized physical-control regimes but does not separate their contributions or establish robustness to any one factor.Source: §IX · p. 9

Limitations

  • The evidence is limited by one checkpoint per learned configuration, restricted temporal coverage, no harmonization ablation, no contemporary learned baseline, and replayed exogenous context.Best-of-six metrics favor probabilistic models over deterministic baselines, and fine-tuning lacks a matched 1,000-window scratch control.Source: §X · p. 9

Results

  • One pooled checkpoint improves both 10-s displacement metrics over local training at all five sites, while leave-one-site-out reuse improves both metrics at four sites.Leave-one-site-out reuse usually sacrifices 2-s accuracy, making the benefit horizon-dependent.Source: §XI · p. 9, §IX · p. 9

APPENDIX A ARTIFACTS

The appendix documents data preparation, shared training settings, fine-tuning, and non-learned longitudinal baselines used in the comparisons.

Data preparation

  • Data is split 80/10/10 by collected-data time, preprocessed separately, and filtered to reduce leakage from trajectories crossing splits.Noisy and incomplete trajectories are removed, and bounding-box endpoints mark movement classes.Source: A. Data Preprocessing · p. 11

Training and fine-tuning

  • The same architecture and hyperparameters are used across all five sites, while fine-tuning samples 1,000 windows from each 100,000-window training split.Fine-tuning updates all model weights; freezing different components is not analyzed.Source: B. Train/Fine-tune · p. 11

Least-squares and IDM baselines

  • The least-squares baseline fits a first-order polynomial to historical ego speed and extrapolates the fitted slope as constant acceleration.Both non-learned predictors are rolled out for the same H steps from the ego’s last observed ground-truth state.Source: C. Least-Squares and IDM Baselines · p. 11
  • IDM sets acceleration from the bumper gap and approach rate, using the nearer forward-cone vehicle or a virtual stop-line leader when the signal is not green.Parameters are calibrated per intersection on the training split using Bayesian posterior means.Source: C. Least-Squares and IDM Baselines · p. 11

Baseline implementation

  • Both baselines project each predicted position onto the ego movement polyline and use the lane tangent as the new heading.This applies the same map constraint to both longitudinal predictors.Source: C. Least-Squares and IDM Baselines · p. 11
Loading 2608.25275v1…