OpenDriveLab/WorldEngine
收藏资源简介：
WorldEngine数据集是一个用于自动驾驶物理人工智能后训练框架的模块化数据集，旨在系统解决自动驾驶中长尾安全关键数据稀缺的问题。该数据集采用模块化数据结构，包括三个主要模块：Raw Data（原始数据，基于nuPlan和OpenScene基础数据集，包含传感器数据、地图和标注）、AlgEngine（算法引擎，用于端到端模型训练和评估，提供预处理标注、检查点和缓存）和SimEngine（模拟引擎，用于闭环仿真环境，提供场景资源和配置文件）。数据集通过数据驱动的长尾发现自动从真实驾驶日志中识别易出错的场景，无需手动设计或合成扰动；利用3D高斯泼溅（3DGS）实现逼真的交互式模拟，将每个发现场景重建为完全可控、实时可渲染的仿真环境；并借助行为世界模型（BWM）从长尾场景中泛化和合成多样化的交通变化，将稀疏的安全关键事件扩展为密集、可学习的分布。该数据集已在量产ADAS平台上进行生产规模验证，基于80,000+小时的驾驶日志训练，能降低碰撞率高达45.5%，并在200公里道路测试中实现零脱离。数据集还包含预训练模型检查点、预计算指标缓存和场景配置等资源，支持自动驾驶系统的后训练和评估。
The WorldEngine dataset is a modular dataset for a post-training framework in Physical AI for autonomous driving, designed to systematically address the long-tail safety-critical data scarcity problem. It features a modular data structure with three main components: Raw Data (based on nuPlan and OpenScene base datasets, including sensor data, maps, and annotations), AlgEngine (for end-to-end model training and evaluation, providing preprocessed annotations, checkpoints, and caches), and SimEngine (for closed-loop simulation environments, offering scene assets and config files). The dataset enables data-driven long-tail discovery, automatically identifying failure-prone scenarios from real-world driving logs without manual design or synthetic perturbations; uses photorealistic interactive simulation via 3D Gaussian Splatting (3DGS) to reconstruct each discovered scenario into a fully controllable, real-time-renderable environment; and leverages Behavior World Model (BWM) to generalize and synthesize diverse traffic variations from long-tail scenarios, expanding sparse safety-critical events into a dense, learnable distribution. It has been production-validated on a mass-produced ADAS platform trained on over 80,000 hours of driving logs, reducing collision rates by up to 45.5% and achieving zero disengagements in a 200 km on-road test. The dataset also includes resources such as pre-trained model checkpoints, pre-computed metric caches, and scenario configurations to support post-training and evaluation of autonomous driving systems.
