DrivingDojo 自动驾驶数据集

DrivingDojo 数据集是由中国科学院自动化研究所模式识别新实验室、中国科学院大学人工智能学院、美团公司和中国科学院香港人工智能与机器人中心于 2024 年联合创建的，相关论文成果为「DrivingDojo Dataset: Advancing Interactive and Knowledge-Enriched Driving World Model」，旨在推进交互式和知识丰富的驾驶世界模型的发展。这个数据集包含约 18k 个视频片段，专门模拟真实世界的视觉交互，涵盖了丰富的驾驶动作、多智能体交互和开放世界的驾驶知识。

DrivingDojo Dataset was jointly created in 2024 by the Pattern Recognition New Laboratory, Institute of Automation, Chinese Academy of Sciences, School of Artificial Intelligence, University of Chinese Academy of Sciences, Meituan, and Hong Kong Institute of Artificial Intelligence and Robotics, Chinese Academy of Sciences. Its associated research paper is titled "DrivingDojo Dataset: Advancing Interactive and Knowledge-Enriched Driving World Model", which is designed to advance the development of interactive and knowledge-enriched driving world models. This dataset contains approximately 18,000 video clips that specifically simulate real-world visual interactions, covering a wide range of driving maneuvers, multi-agent interactions, and open-world driving knowledge.

搜集汇总

背景与挑战

背景概述

DrivingDojo自动驾驶数据集由多个研究机构于2024年联合创建，包含约18k个视频片段，模拟真实驾驶交互，涵盖丰富的驾驶动作、多智能体交互和开放世界知识，视频分辨率为1920×1080、帧率5fps，来自中国多个城市，并配有同步相机姿势和文本描述。该数据集还提出了行动指令跟随基准，旨在提升自动驾驶世界模型的预测和控制能力。