WTR-Drive: World-Task Decomposition with Structured Risk Supervision for End-to-End Driving
Organizations: All authors: Voyah Automotive Technology Co., Ltd., Wuhan, China.
Abstract
Proposal-conditioned scene predictions can lack a common object-level reference across candidate trajectories. Aggregate planning scores also leave the objects and times associated with violations implicit. We present WTR-Drive (World-Task-Risk Driving), a Drive-JEPA-based planner that separates shared scene estimation from proposal-conditioned reasoning. A proposal-independent World Memory estimates object states, map structure, and future motion by re-reading the original visual features at each refinement round. Proposal feedback updates a residual Task branch without writing back to World Memory. This branch guides refinement through candidate-specific retrieval and gated visual feedback. A final Task readout supplies candidate-conditioned features for scoring. We derive structured risk targets from existing candidate-scoring rollouts without additional simulator calls. For at-fault collision and time-to-collision violations, the targets encode event existence, association with shared object slots, and the first violation time bin. On the NAVSIM navtest split, WTR-Drive achieves 94.66 PDMS on v1 and 88.24 EPDMS on v2 (expert-frame sequential evaluation). These scores are 0.96 and 0.44 points above the published Drive-JEPA values, respectively. Candidate-level analysis shows that higher best-candidate scores can coexist with lower selected-trajectory scores, revealing a gap between candidate quality and selection effectiveness.
Figures & tables
| Symbol | Meaning (shape for tensors) |
|---|---|
| Candidates, forecast steps, rounds | |
| Round, candidate, step, object, risk indices | |
| Scene slots, visual tokens, channels | |
| Original / Task-conditioned vision, | |
| Proposal / readout features, | |
| Trajectory coordinates, |
| Method | Backbone | Input | NC | DAC | EP | TTC | C | PDMS |
|---|---|---|---|---|---|---|---|---|
| DiffusionDrive [ 7 ] | ResNet-34 | C+L | 98.20 | 96.20 | 82.20 | 94.70 | 100.00 | 88.10 |
| GoalFlow [ 8 ] | V2-99 | C+L | 98.40 | 98.30 | 85.00 | 94.60 | 100.00 | 90.30 |
| DiffusionDriveV2 [ 28 ] | ResNet-34 | C+L | 98.30 | 97.90 | 87.50 | 94.80 | 99.90 | 91.20 |
| iPad [ 4 ] | ResNet-34 | C | 98.60 | 98.30 | 88.00 | 94.90 | 100.00 | 91.70 |
| DriveSuprim [ 9 ] | ViT-L | C | 98.60 | 98.60 | 91.30 | 95.50 | 100.00 | 93.50 |
| Drive-JEPA [ 3 ] | ViT-L | C | 99.10 | 98.20 | 91.40 | 96.20 | 99.90 | 93.70 |
| Configuration | Readout | PDMS | EPDMS | EPDMS |
|---|---|---|---|---|
| Training configurations | ||||
| Scene-query baseline | – | 94.20 | 88.32 | 91.59 |
| + map refinement | – | 94.43 | 88.45 | 91.84 |
| w/o subscore loss | On | 94.35 | 88.39 | 91.88 |
| + ranking loss | On | 94.22 | 88.35 | 91.75 |
| WTR-Drive | On | 94.66 | 88.24 | 91.55 |
End-to-End Autonomous Driving
End-to-end autonomous driving learns to turn sensor observations and driving context into vehicle actions or trajectories.
See all 253 papers