Orbis 2: A Hierarchical World Model for Driving PredictionOrbis 2：面向自动驾驶预测的分层世界模型
S 1.7T11 sources1 个来源R7-research
On July 17, 2026, Sudhanshu Mittal, Thomas Brox and five coauthors posted "Orbis 2: A Hierarchical World Model for Driving" (arXiv:2607.15898v1), arguing that existing world models operate at a single level of abstraction and mostly prioritize perceptual fidelity while lacking the spatial reasoning and semantic understanding needed for real-world downstream tasks.
Orbis 2 factorizes future prediction into two levels at distinct temporal and abstraction scales: a high-level predictor that forecasts coarse scene structure over extended temporal horizons, and a low-level generator that produces detailed predictions conditioned on that high-level output.
By splitting long-horizon scene structure from fine-grained detail generation, the hierarchical design targets the spatial-reasoning and semantic-understanding gap the authors identify in current single-level driving world models, a capability directly relevant to downstream autonomous-driving planning and decision tasks.
年7月17日，Sudhanshu Mittal、Thomas Brox 等7位作者在arXiv发布论文《Orbis 2: A Hierarchical World Model for Driving》（arXiv:2607.15898v1），指出现有世界模型多停留在单一抽象层级，偏重感知保真度，却缺乏真实下游任务所需的空间推理与语义理解能力。
Orbis 2将未来预测拆分为两个不同时间与抽象尺度的层级：高层预测器负责在较长时间跨度上预测粗粒度场景结构，低层生成器则在高层输出的条件下生成细节丰富的预测结果。