Enabling a Novel Evaluation Continuum for Connected & Autonomous Vehicles (CAV)

全球互联和自动驾驶汽车（CAV）行业估计价值超过500亿英镑（到2035年），英国CAV行业预计超过30亿英镑。此外，英国政府的工业战略旨在到2021年在英国道路上实现完全自动驾驶的汽车，而英国是世界上首批实现这一目标的国家之一。然而，为了实现这一愿景和市场潜力，安全引入CAV是必要的，需要大量的研究，以克服与CAV的公共部署相关的各种障碍（技术，立法和社会）。虽然原型CAV技术已经存在了一段时间，但确保这些技术的安全水平已被证明是CAV技术商业化的障碍。CAV的愿景与测试和安全分析的挑战相结合，因为它需要复杂的解决方案，包括大量变量和环境之间的相互作用。据建议，为了证明CAV比人类驾驶员更安全，他们将需要驾驶超过110亿英里。该奖学金的愿景是支持将英国定位为CAV研究和创新的世界领导者，以实现长期的社会和经济效益。该奖学金将开发开创性的测试方法和标准，以实现CAV的稳健和安全使用，重点是创建基础知识和应用研究方法和工具。在英国沃里克大学的WMG，我们为CAV创建了“评估连续体”的概念，包括使用各种环境，如数字世界，模拟环境，测试跟踪测试和真实世界的测试。有两个方面是共同的每一个评估连续环境，也是奖学金研究的重点领域1）测试场景（环境输入）2）安全证据（环境输出）。关于场景主题，虽然110亿英里的要求已经获得了很多宣传，但重点需要放在这些英里中发生的事情上（即，智能英里，暴露CAV中的故障），而不是英里数本身。作为该奖学金的一部分，将探讨三种方法来确定这些智能里程。这些包括1）使用基于机器学习（ML）的方法，包括贝叶斯优化，为测试场景创建测试用例，2）预期功能的安全性（SOTIF）（创新的CAV安全分析）为基础的测试场景和3）将真实世界的数据转换为可执行的测试场景的仿真工具。所有这些方法将共同有助于创建英国的国家CAV测试场景数据库，这将有助于协调由英国政府部分资助的各种CAV项目中的研究工作，并将防止每个项目中关于测试场景识别的“重新发明轮子”。CAV的行业趋势表明机器学习（ML）在自主控制系统中的广泛采用。ML系统的结构本质上是不确定的，这使得CAV系统本质上是高度不透明的。因此，很难识别这种基于ML的系统中的故障原因并采取纠正措施。因此，在安全链上，基础研究将作为该奖学金的一部分进行，以探索如何使基于ML的系统可解释，使我们能够解释结果。由于CAV部署的关键性质和风险的减轻，这是CAV安全的基本要求。此外，该奖学金还将受益于该研究员作为英国ISO标准委员会技术代表的第一手经验，提供进一步的见解和明确的路线，通过制定国际标准，同时也确保英国成为这一领域的世界领导者。

The global Connected & Autonomous Vehicles (CAV) industry is estimated to be worth over £50billion (by 2035), with the UK CAV industry being projected over £3billion. Additionally, the UK Government's Industrial Strategy aims to bring fully autonomous cars without a human operator on the UK roads by 2021, one of the first countries in the world to achieve this. However, in order to realise this vision and the market potential, safe introduction of CAV is necessary, requiring significant research to overcome diverse barriers (technological, legislative and societal) associated with public deployment of CAV.While prototype CAV technologies have existed for some time now, ensuring the safety level of these technologies has been proving to be a hindrance to the commercialization of CAV technologies. The vision for CAV is coupled with the challenge of testing and safety analysis as it needs complex solutions to include interactions between a large number of variables and the environment. It is suggested that in order to prove that CAV are safer than human drivers, they will need to be driven for more than 11 billion miles.The vision for this fellowship is to support positioning the UK as the world leader in CAV research and innovation for a long lasting societal and economic benefit. This fellowship will develop pioneering testing methodologies and standards to enable robust and safe use of CAV with a focus on creating both fundamental knowledge and applied research methods and tools. At WMG, University of Warwick, UK, we have created a concept of the "evaluation continuum" for CAV, which involves using various environment like digital world, simulated environment, test track testing and real-world for testing.There are two aspects which are common to each of the evaluation continuum environments and also the focus areas of the fellowship research 1) Test Scenarios (input to the environment) 2) Safety Evidence (output of the environment). On the scenarios theme, while the 11-billion-miles requirement has garnered a lot of publicity, the focus needs to be on what happens in those miles (i.e., smart miles which expose failures in CAV) and not on the number of miles themselves. As a part of this fellowship, three approaches will be explored to identify these smart miles. These include 1) using Machine Learning (ML) based methods including Bayesian Optimisation to create test cases for test scenarios, 2) Safety Of The Intended Functionality (SOTIF) (Innovative safety analysis of CAV) based test scenarios and 3) translating real-world data into executable test scenarios for a simulation tool.All these approaches will together contribute to the creation of a UK's National CAV Test Scenario Database, which will help coordinate the research work in various CAV projects part-funded by the UK Government and will prevent "reinventing of the wheel" in each of the projects with respect to test scenario identification.Industry trends in CAV suggest the widespread adoption of machine learning (ML) in the autonomous control systems. ML-systems by their structure are non-deterministic in nature, making the CAV system highly opaque in nature. Therefore, it is difficult to identify the reason of a failure in such ML-based systems and take the corrective measures. Thus, on the safety strand, fundamental research will be conducted as a part of this fellowship to explore how to make ML-based systems interpretable enabling us to explain the results. This is an essential requirement for safety of CAV due to the critical nature of their deployment and the mitigation of risk.In addition, the fellowship will also benefit from the fellow's first-hand experience as the UK's technical representative on the ISO standards committees, providing further insight and a clear route to deliver impact from the proposed research through the development of international standards, while also ensuring that the UK becomes a world leader in this area.