Driving · 2026-10-14
MORE: Learning Motion-Oriented Registers for Driving World Models and Planning
Yingdong Gu, Haoyu Wang, Shaocheng Yan, Zhenjun Zhao, Hao Han, Pengcheng Shi, Xianfei Li, Qinghua Tan, Wenlong Liao, Tao He, Jiayuan Li, Pai Peng
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
MORE learns motion-oriented registers during video pretraining (a V-JEPA framework) and uses the same compact register state interface for both a non-autoregressive driving world model and trajectory planning. On NAVSIM-v2, the authors report the highest EPDMS (92.5), and in efficiency analysis they report 1/10 the inference latency of DriveLaW while keeping inference latency close to Drive-JEPA (110.0 vs 73.3 ms for Drive-JEPA).
Why it matters
The authors argue that driving world models/planners require ego-motion-sensitive future states, but existing methods inherit dense visual features and adapt them downstream. MORE instead shapes the downstream state interface directly during video pretraining via motion-conditioned, register-bottlenecked masked prediction, then reuses that register space for future-state modeling and shared scene memory for planning.
Method
- V-JEPA-style video pretraining with a small set of learnable motion-oriented registers; ego-motion is injected into registers and they are the sole context for masked patch prediction (register-mediated masked patch prediction).
- A complementary blockwise register-masking objective masks contiguous blocks of registers and predicts masked register states, shaping predictive structure in the register space.
- Joint training reuses the same register space: a non-autoregressive world model predicts future registers from observed registers and ego motion, and a planner uses observed+predicted registers as shared scene memory to generate and score K candidate trajectories.
Abstract (from arXiv)
By predicting future states, latent world models provide rich self-supervision for end-to-end driving, but often inherit dense features from pretrained visual encoders as their world-state interface. This creates a mismatch: visual pretraining produces information-rich, spatially dense features, but their adaptation for motion-conditioned world modeling and planning is deferred to downstream training. The dense grid also makes future prediction more computationally expensive as visual resolution increases. We therefore argue that video pretraining should directly learn state representations suited to downstream world modeling and planning, rather than relying on downstream training to adapt inherited visual features. We propose ***MORE***, a V-JEPA framework that learns **M**otion-**O**riented **RE**gisters during video pretraining as a shared state interface for driving world models and planning. MORE injects temporally aligned ego motion into a small set of registers and uses them as the sole context for masked patch prediction, channeling dense patch-level supervision through a compact state bottleneck. A complementary blockwise register-masking objective further shapes predictive structure in the register space. During downstream training, the same register space defines the future states predicted by the world model and provides shared scene memory for trajectory generation and scoring. On NAVSIM-v2, MORE outperforms the state of the art by 1.6 in EPDMS, while achieving superior efficiency with only 1/10 the inference latency of DriveLaW. These results support learning the state interface for world modeling and planning directly during video pretraining.
Related papers
- Pretrained Features as Teachers, Not States: Compact Latent World Models for Visual Planning
- PlanWAM: Planning-Shaped Future Representations for End-to-End Autonomous Driving
- Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction
- VGGTWorld-VLA: Intent-Conditioned 3D World Evolution for Autonomous Driving
- SkipWorld: Latent Skip Planning with Distilled World Models for Efficient Visual Navigation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-10 · Indexed by Pub-AI: 2026-10-14 UTC
MORE: Learning Motion-Oriented Registers for Driving World Models and Planning
Preprint posted: 2026-10-10 · Original record ↗
- arxiv
- 2610.14871
Explicit source versions (1)
- v1 · 2026-10-10
- Preprint available
Source date: 2026-10-10 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
By predicting future states, latent world models provide rich self-supervision for end-to-end driving, but often inherit dense features from pretrained visual encoders as their world-state interface. This creates a mismatch: visual pretraining produces information-rich, spatially dense features, but their adaptation for motion-conditioned world modeling and planning is deferred to downstream training. The dense grid also makes future prediction more computationally expensive as visual resolution increases. We therefore argue that video pretraining should directly learn state representations suited to downstream world modeling and planning, rather than relying on downstream training to adapt inherited visual features. We propose ***MORE***, a V-JEPA framework that learns **M**otion-**O**riented **RE**gisters during video pretraining as a shared state interface for driving world models and planning. MORE injects temporally aligned ego motion into a small set of registers and uses them as the sole context for masked patch prediction, channeling dense patch-level supervision through a compact state bottleneck. A complementary blockwise register-masking objective further shapes predictive structure in the register space. During downstream training, the same register space defines the future states predicted by the world model and provides shared scene memory for trajectory generation and scoring. On NAVSIM-v2, MORE outperforms the state of the art by 1.6 in EPDMS, while achieving superior efficiency with only 1/10 the inference latency of DriveLaW. These results support learning the state interface for world modeling and planning directly during video pretraining.