Collaborative Research: CCRI: New: An Open Source Simulation Platform for AI Research on Autonomous Driving

合作研究：CCRI：新：自动驾驶人工智能研究的开源仿真平台

• 批准号: 2235012
• 负责人: Bolei Zhou
• 金额: $ 96.03万
• 依托单位: University of California-Los Angeles
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-03-15 至 2026-02-28
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2235012&HistoricalAwards=false
• 关键词: Collaborative; Research; CCRI; New; Open

Autonomous driving is transforming daily life and economy, with promised benefits like safe transportation and efficient mobility. Much of today’s research on autonomous driving experiments on expensive commercial vehicles. It is costly and risky to evaluate AI and machine learning methods on physical vehicles in real world. Driving simulator provides a cost-effective and safe alternative for the development and evaluation of new AI algorithms. However, existing driving simulators with limited assets and complexity cannot accommodate the needs of the rapidly progressing AI fields. This project aims to develop an open-ended driving simulation platform that fosters innovations in various aspects of autonomous driving from perception to decision-making. This platform will support realistic driving simulation with a diverse range of traffic assets and scenarios imported from real world. It will become a common experimental ground for researchers in academia and industry to develop new AI methods, share data and models, and benchmark the progress. The platform will grow into a community research infrastructure and have significant impacts on the blooming autonomous driving industry. Additionally, it will provide interactive teaching toolkits for STEM education, particularly for students from underserved communities. In this project, investigators will develop an open-source simulation platform called MetaDriverse for AI research on autonomous driving. This platform will serve as a research infrastructure and facilitate compelling research opportunities in various disciplines, including computer vision, computer graphics, machine learning, and human-machine interaction. MetaDriverse will feature realistic visual appearance, interactive real-world scenarios and assets, and intuitive human control interface, allowing for the simulation of real-world driving experiences. It will also provide a wide range of tasks and benchmarks and the flexibility to design new ones, which will gauge the community’s collective effort and accelerate the research progress. The key features and capabilities of the platform include (i) realistic visual perception and neural rendering, (ii) interactive simulation of real-world traffic scenarios, (iii) comprehensive benchmarks and model zoo, and (iv) community’s collective effort. This infrastructure will foster a wide range of opportunities and collaborations in the autonomous driving and intelligent transportation industries and bring significant societal and economic impacts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

自动驾驶正在改变日常生活和经济，并带来安全运输和高效流动等好处。 今天的大部分研究都是在昂贵的商用车上进行自动驾驶实验。在真实的世界中的物理车辆上评估人工智能和机器学习方法是昂贵和危险的。驾驶模拟器为开发和评估新的AI算法提供了一种具有成本效益和安全的替代方案。然而，现有的驾驶模拟器具有有限的资产和复杂性，无法满足快速发展的人工智能领域的需求。该项目旨在开发一个开放式驾驶模拟平台，促进从感知到决策的自动驾驶各个方面的创新。该平台将支持逼真的驾驶模拟，并从真实的世界导入各种交通资产和场景。它将成为学术界和工业界研究人员开发新的人工智能方法，共享数据和模型并对进展进行基准测试的共同实验场。该平台将发展成为一个社区研究基础设施，并对蓬勃发展的自动驾驶行业产生重大影响。 此外，它还将为STEM教育提供互动式教学工具包，特别是针对来自服务不足社区的学生。在这个项目中，研究人员将开发一个名为MetaDriverse的开源模拟平台，用于自动驾驶的人工智能研究。该平台将作为研究基础设施，促进各个学科的引人注目的研究机会，包括计算机视觉，计算机图形学，机器学习和人机交互。MetaDriverse将具有逼真的视觉外观，交互式现实世界场景和资产，以及直观的人类控制界面，允许模拟现实世界的驾驶体验。它还将提供广泛的任务和基准，并灵活地设计新的任务和基准，这将衡量社区的集体努力，并加快研究进展。该平台的主要特点和功能包括（i）逼真的视觉感知和神经渲染，（ii）真实世界交通场景的交互式模拟，（iii）全面的基准和模型动物园，以及（iv）社区的集体努力。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

ScenarioNet: Open-Source Platform for Large-Scale Traffic Scenario Simulation and Modeling

• DOI: 10.48550/arxiv.2306.12241
• 发表时间: 2023-06
• 期刊: ArXiv
• 作者: Quanyi Li;Zhenghao Peng;Lan Feng;Zhizheng Liu;Chenda Duan;Wen-An Mo;Bolei Zhou

CAT: Closed-loop Adversarial Training for Safe End-to-End Driving

• DOI: 10.48550/arxiv.2310.12432
• 发表时间: 2023-10
• 期刊: ArXiv
• 作者: Linrui Zhang;Zhenghao Peng;Quanyi Li;Bolei Zhou

Learning from Active Human Involvement through Proxy Value Propagation

• 发表时间: 2023
• 作者: Zhenghao Peng;Wenjie Mo;Chenda Duan;Quanyi Li;Bolei Zhou