World-model-driven multi-target distilled evaluation for end-to-end autonomous driving
Measurement Science and Technology · 03.07.2026 · 37 (27) · 275105
DOI: 10.1088/1361-6501/ae80ef ↗Kurz gefasst
Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 06.07.2026. Verfasst von Ziheng Lu und 4 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: Jiangsu Province Frontier Technology Research and Development Plan, National Natural Science Foundation of China.
Abstract (Originalsprache)
Sensor-rich end-to-end autonomous driving increasingly relies on auxiliary scorer supervision to compensate for the limitations of pure imitation learning in safety-critical decision-making scenarios. However, most existing trajectory scorers still evaluate candidates only under the current observation, which makes it difficult to assess how a present action may influence future interactions in dynamic traffic scenes. To address this issue, we propose WoEval, a world-model-driven multi-target distilled evaluation framework for end-to-end planning. WoEval introduces a conditional bird’s-eye-view world model to roll out candidate-specific future scene evolution and uses the predicted scene sequence to evaluate candidate trajectories, shifting supervision from static rule matching to future-aware behavior quality assessment. To improve transferability, we decouple trajectory evaluation from candidate generation and backbone-specific planning modules, allowing the evaluator to serve as a reusable scoring component across heterogeneous planners. We further design a structured reward model with high-order scene-trajectory interaction aggregation and combine human-teacher imitation signals with world-model-teacher rule signals for multi-target distillation. Experiments on the NAVSIM benchmark show that WoEval improves candidate ranking quality and yields stronger closed-loop planning performance, while the decoupled design remains compatible with different planning backbones.
Förderung
- Jiangsu Province Frontier Technology Research and Development Plan
- National Natural Science Foundation of China
Zitieren
APA
Lu, Z., Cai, Y., Dong, W., Wang, H., & Chen, L. (2026). World-model-driven multi-target distilled evaluation for end-to-end autonomous driving. Measurement Science and Technology, 37(27), 275105. https://doi.org/10.1088/1361-6501/ae80ef
Vancouver
Lu Z, Cai Y, Dong W, Wang H, Chen L. World-model-driven multi-target distilled evaluation for end-to-end autonomous driving. Measurement Science and Technology. 2026;37(27):275105. doi:10.1088/1361-6501/ae80ef
BibTeX
@article{Lu2026,
author = {Lu, Ziheng and Cai, Yingfeng and Dong, Wei and Wang, Hai and Chen, Long},
title = {World-model-driven multi-target distilled evaluation for end-to-end autonomous driving},
journal = {Measurement Science and Technology},
year = {2026},
volume = {37},
number = {27},
pages = {275105},
doi = {10.1088/1361-6501/ae80ef}
}
EUVIMED