OpenDriveLab/SimScale

SimScale是一个大规模的自动驾驶模拟数据集，旨在通过高保真度的反应式驾驶场景模拟和伪专家演示来支持自动驾驶技术的研究和开发。数据集提供了一个可扩展的模拟管道，能够合成多样化和高保真度的反应式驾驶场景，并包含有效的模拟-现实协同训练策略，以提高各种端到端规划器的鲁棒性和泛化能力。此外，数据集还揭示了模拟-现实学习系统在端到端自治中的关键扩展属性。数据集基于nuPlan和NAVSIM，包含大量的模拟数据，分为规划器基础的伪专家和基于恢复的伪专家两大类，每类又包含多个子集。

SimScale is a large-scale autonomous driving simulation dataset designed to support research and development in autonomous driving technology through high-fidelity reactive driving scenario simulations and pseudo-expert demonstrations. The dataset provides a scalable simulation pipeline capable of synthesizing diverse and high-fidelity reactive driving scenarios, along with an effective sim-real co-training strategy to enhance the robustness and generalization of various end-to-end planners. Additionally, the dataset reveals crucial scaling properties of sim-real learning systems for end-to-end autonomy. Based on nuPlan and NAVSIM, the dataset includes extensive simulation data, categorized into planner-based pseudo-experts and recovery-based pseudo-experts, each containing multiple subsets.