面向多风格端到端驾驶的长时程一致交互感知世界模型

Original title: Long-Horizon Consistent and Interaction-Aware World Models for Multi-Style End-to-End Driving

Source: arXiv cs.RO·2026-09-04 12:00Publish time unverified

arXiv:2609.03225v1 Announce Type: new Abstract: End-to-end autonomous driving has increasingly adopted world model-based reinforcement learning frameworks to improve learning efficiency through \textit{imagined rollouts}. However, existing world models suffer from three key limitations: temporal inconsistency in long-horizon imagined rollouts, inadequate modeling of ego-environment interactions, and limited adaptability to diverse driving styles.