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.
Figures & tables
Fig. 1: PhaseShift’s topology-aware multi-intersection pretraining and deployment workflow. A shared spatial–temporal backbone is trained across harmonized domains and reused as one pooled model, deployed zero-shot at a new intersection, or fine-tuned where local data show that specialization is needed. The domain-diagnosis path treats residual zero-shot error as a prioritization signal, not as proof of a particular behavioral cause.
Site
Train (min)
Flow (veh/h)
Movements
Cycle (s)
Cycle CoV
Mean green split (%)
7
48
2672.93
16
150.0
0.1
19.6
8
48
1126.07
11
131.0
0.3
27.8
9
56
985.37
8
120.0
0.3
46.0
10
84
1332.15
14
117.0
0.3
27.6
22
19
5061.54
24
179.0
0.0
24.0
TABLE I: Characteristics of the five intersection domains. “Train” is the duration of recorded traffic from which each site’s 100,000 training samples are drawn; the total collected recordings are longer (0.42–1.75 h per site). Flow is the rate of vehicles assigned to a movement cluster (clustering uses each track’s starting and ending bounding boxes; vehicles missing either endpoint are excluded); Movements is the number of induced movement clusters; CoV is the coefficient of variation of the signal cycle length; green split is the mean green fraction.
Fig. 2: The five intersection domains used by PhaseShift. The panels are cropped around the monitored conflict zones and are not shown at a common scale.
Target
Method
NLL (2 s)
minADE (2 s)
minFDE (2 s)
NLL (5 s)
minADE (5 s)
minFDE (5 s)
NLL (10 s)
minADE (10 s)
minFDE (10 s)
7
Const.-Vel.
–
1.8559
3.6330
–
2.0404
4.2413
–
2.0216
4.3092
7
IDM
–
1.8021
3.4662
–
1.8026
3.8486
–
1.7216
4.0604
7
Local
-4.6159
0.2800
0.5138
-4.2835
0.8262
1.8892
-4.0209
2.1595
5.1121
7
Zero-Shot
-4.5001
0.4039
0.7651
-5.2363
0.7449
1.9011
-5.6398
0.9247
2.8034
7
Fine-Tuned
-5.0538
0.3577
0.6388
-5.6513
0.6523
1.6335
-5.9421
0.9263
2.9048
7
Pooled
-5.1002
0.2402
0.4105
-5.3704
0.4781
1.1053
-5.4502
0.7111
2.0438
TABLE II: Multi-intersection test results. NLL is in nats; minADE/minFDE are in meters over N=6 sampled rollouts. Lower is better. Bold marks the best single-run displacement value per site and the best learned-model NLL. Macro averages include all five sites.
Fig. 3: Domain-level error reduction relative to independently trained Local models. Positive values mean the shared-backbone regime has lower error; negative values indicate a residual specialization gap. “Mean” and “Med.” summarize site-level percentage reductions.
Split 7/8/9/10/22
minADE (2 s)
minFDE (2 s)
minADE (5 s)
minFDE (5 s)
minADE (10 s)
minFDE (10 s)
100/0/0/0/0
0.2800
0.5138
0.8262
1.8892
2.1595
5.1121
80/5/5/5/5
0.2492
0.4268
0.4774
1.2210
0.7309
2.1264
60/10/10/10/10
0.2241
0.3843
0.4912
1.2340
0.8315
2.5654
40/15/15/15/15
0.2599
0.4607
0.5560
1.2933
1.0697
2.9718
20/20/20/20/20
0.2438
0.4015
0.4641
1.0681
0.9986
2.8753
TABLE III: Fixed-100,000-window composition sweep at site 7. Splits give the percentage drawn from sites 7/8/9/10/22; 100/0/0/0/0 is Local . Bold marks the best value per column. One run per composition.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Scene context
Time step Δt
0.1s
Observation history Th
20 steps ( 2.0s )
Prediction horizon Tf
20 steps ( 2.0s )
Max. neighbors
10
Neighbor radius
15m
Appendix
TABLE IV: Model and training hyperparameters.
Parameter
Pre-training
Fine-tuning
Learning rate
1×10−3
1×10−5
Batch size
256
64
Warmup steps
4000
20
Training windows M
100,000
1,000
Appendix
TABLE V: Fine-tuning hyperparameters. All values not listed are inherited unchanged from the pre-trained model (Table IV ).
School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong · Shanghai AI Laboratory, Shanghai, China +1