WZDrive
Human-in-the-loop driving data for
trajectory prediction in roadway work zones
Demo
Work-zone scenes, recorded signals, and trajectory prediction.
Dataset at a glance
Driving-trajectory data from roadway work zones remain scarce.
Safety-critical encounters cannot be staged safely on real roads, so we use a human-in-the-loop simulation.
Three work-zone configurations across six scenario conditions.
nuScenes: 1,000 distinct 20-second driving scenes (≈5.5 h), each with an ego-vehicle trajectory.
Data collection platform: DuoRoad
Results
Two-second observation and five-second prediction at 5 Hz. Errors are in meters; lower is better.
Rule-based & time-series baselines
| Model | Features | ADE (m) ↓ | FDE (m) ↓ |
|---|---|---|---|
| Stay | Ego | 21.635 | 42.053 |
| CV | Ego | 7.237 | 16.896 |
| LSTM | Ego + control + WZ | 2.988 | 7.513 |
| Transformer | Ego + control + WZ | 2.876 | 7.284 |
| GRU | Ego + control + WZ + lane | 2.860 | 7.320 |
One predicted trajectory (K = 1). WZ denotes work-zone context.
SOTA baselines
| Model | Without WZ | With WZ | ||
|---|---|---|---|---|
| minADE6 ↓ | minFDE6 ↓ | minADE6 ↓ | minFDE6 ↓ | |
| PGP (2021) | 1.286 | 2.749 | 1.260↓2.02% | 2.742↓0.25% |
| MTR (2022) | 1.317 | 2.660 | 1.299↓1.37% | 2.567↓3.50% |
| Forecast-MAE (2023) | 1.373 | 2.650 | 1.264↓7.94% | 2.475↓6.60% |
| LAFormer (2024) | 1.204 | 2.662 | 1.179↓2.08% | 2.550↓4.21% |
| DeMo (2024) | 1.514 | 2.436 | 1.281↓15.39% | 2.462↑1.07% |
| RouteFormer (2025) | 2.106 | 4.023 | 2.050↓2.65% | 3.847↓4.38% |
| MM-GRU (ours)† | 1.234 | 2.682 | 1.124↓8.91% | 2.242↓16.41% |
| WZDrive (ours)* | — | — | 1.094 ± 0.022↓2.67% | 2.115 ± 0.082↓5.66% |
Six predicted trajectories (K = 6). With WZ includes temporary work-zone geometry and static worker-position context; Without WZ omits those inputs. WZDrive values are mean ± SD across three seeds.
For each model from PGP through MM-GRU, percentages compare With WZ against that model's Without WZ result. WZDrive* compares with MM-GRU with WZ. ↓ means lower error; ↑ means higher error.
† MM-GRU is our best-performing GRU variant among the multimodal input combinations we tested. It predicts six trajectories using ego motion, driver controls, surrounding agents, lane geometry, and work-zone context.
* WZDrive is our final model, using modality-specific fusion, staged training, direct six-path decoding, and temporal calibration. Only its work-zone-aware configuration is reported here.