EditWM: Event-Decomposed World Modeling with Incremental Correction for End-to-End Autonomous Driving

Junjie Yang, Qingwei Zeng, Youyou Li, Zicheng Ding, Ziyi Shi, Shuqi Shen, Hongliang Lu, Hai Yang · N/A · 2026

Summary

Abstract Summary

World models support autonomous driving by predicting the scene evolution associated with candidate trajectories. Driving dynamics differ in predictability, motivating a distinction between regular evolution and event-induced deviations that call for selective correction. We propose EditWM, a world model that decomposes future prediction into normal evolution and event-driven incremental correction in compact visual feature space. A trajectory-conditioned normal predictor provides the base forecast and is then frozen for correction learning. A correction decoder compares this forecast with observation history and planned actions, producing a bounded feature update whose contribution is regulated by a learned gate. The corrected future features condition trajectory scoring through candidate-specific cross-attention, linking world modeling to plan selection. At inference, EditWM uses only past and current observations, ego state, and candidate trajectories. Across all 12,146 NAVSIM navtest scenes, expert-trajectory-conditioned evaluation shows a 5.35\% reduction in future-feature MSE over Normal, with improvements in 83.54\% of scenes. The system achieves 91.05 EPDMS on a 100-point scale using the official EPDMS evaluator. These results demonstrate improved future-feature prediction and competitive trajectory selection when corrected future representations are integrated into planning.

Key Points

World models support autonomous driving by predicting the scene evolution associated with candidate trajectories.
Driving dynamics differ in predictability, motivating a distinction between regular evolution and event-induced deviations that call for selective correction.
We propose EditWM, a world model that decomposes future prediction into normal evolution and event-driven incremental correction in compact visual feature space.
A trajectory-conditioned normal predictor provides the base forecast and is then frozen for correction learning.
A correction decoder compares this forecast with observation history and planned actions, producing a bounded feature update whose contribution is regulated by a learned gate.

Method

At time $t$, a frozen DINOv2-based encoder [ 23 ] maps multi-view images to history $\mathcal{H}_{t}=\{Z_{t-L+1},\ldots,Z_{t}\}$. Each of the $L$ frames contains $N$ visual tokens of dimension $d$, giving $Z\in\mathbb{R}^{N\times d}$.

The planner also receives ego state $e_{t}$ and candidates $\mathcal{C}_{t}=\{\tau_{i}\}_{i=1}^{K}$, where $K$ is the candidate count and $\tau_{i}\in\mathbb{R}^{H\times 3}$ contains $H$ time-indexed planar positions and headings in the current ego frame. As shown in Fig.

Method — EditWM: Event-Decomposed World Modeling with Incremental Correction for End-to-End Autonomous Driving — Fig. 4: Selected qualitative planning comparisons using previously evaluated checkpoints. Each row presents one scene, with a camera view followed by Ours, Normal-only, and WoTE. The black dotted line is the human GT trajectory; colored lines show the actual selected trajectories. BEV panels share the same map, current objects, and spatial extent within each row. EPDMS and its displayed sub-scores use a 0–100 scale. GT ADE is the mean position error over eight corresponding future time steps. Future MSE measures expert-trajectory-conditioned future DINO feature prediction for EditWM and Normal.

Experiments

∗ Original v2 evaluator before human-behavior filtering correction. Baseline results are taken from their reported evaluations. Best displayed values, including ties, are bold. We use NAVSIM [ 30 ] , training the trajectory scorer on all 103,288 navtrain scenes and evaluating on all 12,146 navtest scenes.

World-model training uses the 43,849 navtrain scenes with real future images, divided by source log into 39,337 fitting and 4,512 validation scenes. These world-model splits and navtest do not share physical frames.

Sources and Demonstrations

Figures are reproduced from Yang et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.