用鸟瞰图+语言模型，让自动驾驶更准更安全
BEV-VLM: Trajectory Planning via Unified BEV Abstraction
- 用相机与激光雷达融合的鸟瞰图作为输入
- 在nuScenes上准确率提升53.1%，零碰撞
- 适合想用大模型做规划的自动驾驶研究者
本文提出BEV-VLM，一种基于视觉-语言模型（VLMs）的自动驾驶轨迹规划新方法。该方法以融合相机与激光雷达数据生成的鸟瞰图（BEV）特征图为视觉输入，并与高精地图（HD map）对齐，形成几何一致且语义丰富的统一表示。相比仅依赖原始图像的现有方法，该框架显著提升规划精度，在nuScenes数据集上的实验表明，规划准确率提升53.1%，评估场景中实现完全避障。结果证明，VLM可有效理解如BEV等处理后的视觉表征，拓展其在轨迹规划任务中的应用范围。
原文摘要 · Abstract (English)
This paper introduces BEV-VLM, a novel approach for trajectory planning in autonomous driving that leverages Vision-Language Models (VLMs) with Bird's-Eye View (BEV) feature maps as visual input. Unlike conventional trajectory planning approaches that rely solely on raw visual data (e.g., camera images), our method utilizes a highly compressed and informative BEV representation generated by fusing camera and LiDAR data, with subsequent alignment to High-Definition (HD) maps. This unified BEV-HD map format provides a geometrically consistent and semantically rich scene description, which enables VLMs to perform accurate and robust trajectory planning. Experimental results on the nuScenes dataset demonstrate that, compared with state-of-the-art vision-only methods, our approach achieves a 53.1% improvement in planning accuracy and realizes complete collision avoidance in evaluation scenarios. Our work highlights that VLMs can effectively interpret processed visual representations such as BEV features, expanding their applicability beyond raw image inputs for the task of trajectory planning.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。