中文利用集成扩散估计端到端自动驾驶的不确定性

EnDfuser提出一种基于扩散模型的端到端驾驶规划系统，有效融合相机与LiDAR感知信息，通过注意力机制处理不确定性。相比不考虑或依赖专用表示的方法，该方法在CARLA仿真中表现更优，具有更好的通用性和实际部署潜力。

arXiv:2506.00560v2 Announce Type: replace-cross Abstract: End-to-end planning systems for autonomous driving are rapidly improving, especially in closed-loop simulation environments like CARLA. Many such driving systems either do not consider uncertainty as part of the plan itself or obtain it by using specialized representations that do not generalize. In this paper, we propose EnDfuser, an end-to-end driving system that uses a diffusion model as the trajectory planner. EnDfuser effectively leverages complex perception information like fused camera and LiDAR features, through combining attent