UKRI Trustworthy Autonomous Systems Node in Verifiability

自动驾驶系统承诺改善我们的生活；无人驾驶火车和机器人清洁工就是自主系统的例子，它们已经存在于我们中间，并在有限的环境中很好地工作。现在是我们努力确保开发人员能够为动态环境设计可信赖的自主系统并提供其可信性的证据的时候了。由于自主系统的复杂性，通常涉及人工智能组件、低级别硬件控制以及与人类和不确定环境的复杂交互，任何性质的证据都需要各种学科的努力。为了应对这一挑战，我们召集了人工智能、机器人学、人机交互、系统和软件工程以及测试方面的专家联盟。我们将共同为自主系统的属性验证建立基础和技术，以进行信息设计，提供关键属性的证据，并指导部署后的监控。目前，可验证性受到以下几个问题的阻碍：难以理解关注系统个别方面的技术(例如，控制工程、人工智能或人机交互)提供的证据如何构成系统整体的证据；在其学科中使用不同语言和实践的利益相关者之间的沟通困难；处理人工智能、控制和硬件设计、关键系统的软件等方面的高级概念方面的困难。因此，自主系统通常是使用先进的工程技术开发的，但验证方法过时。我们提出了一个创造性的工作计划，将使目前的技术和实践状态能够发生根本变化。我们将定义一个数学框架，使人们能够对自主性核查中涉及的各种做法和概念有一个共同的理解。我们的框架将提供任何工程工作所需的数学基础，以适应不同学科使用的符号。有了这个共同的理解，我们将证明语言之间的翻译，以不同语言定义的人工产物(工程模型、测试、模拟等)的组成，以及从组件验证的系统级推断。有了如此丰富的基础和丰富的结果，我们将改变实践状态。目前，开发人员在没有任何运行条件的证据的情况下，从头开始构建系统，或者重用组件。所产生的系统被部署在受限的条件下(例如，速度降低或受限制的环境)，或者被提供用于部署，用户自负风险。相反，我们设想的是一个经过验证的自主系统和组件商店的未来可用性。在这样的未来，在商店里，用户不仅可以找到系统实现，还可以找到他们操作条件和预期行为的证据(工程模型、数学结果、测试等)。当开发人员签入产品时，商店将需要所有用易于理解的语言描述的搜索结果，并将自动验证可信度的证据。开发人员还将能够为其他开发人员签入组件；同样，他们将获得对其使用的信心所需的证据。在这个变化的世界中，用户将购买具有明确的操作要求和配置文件保证的应用程序。例如，用户还可以要求验证定制平台和环境的充分性。验证不再是问题。与EPSRC TAS集线器和其他节点，以及我们广泛的医疗和工业合作伙伴合作，我们将合作，以确保我们认为的名称、验证技术和属性有助于我们为日常生活带来自治的共同议程。

Autonomous systems promise to improve our lives; driverless trains and robotic cleaners are examples of autonomous systems that are already among us and work well within confined environments. It is time we work to ensure developers can design trustworthy autonomous systems for dynamic environments and provide evidence of their trustworthiness.Due to the complexity of autonomous systems, typically involving AIcomponents, low-level hardware control, and sophisticated interactions withhumans and the uncertain environment, evidence of any nature requires efforts from avariety of disciplines. To tackle this challenge, we gathered consortium ofexperts on AI, robotics, human-computer interaction, systems and softwareengineering, and testing. Together, we will establish the foundations andtechniques for verification of properties of autonomous systems to informdesigns, provide evidence of key properties, and guide monitoringafter deployment.Currently, verifiability is hampered by several issues: difficulties tounderstand how evidence provided by techniques that focus on individualaspects of a system (control engineering, AI, or humaninteraction, for example) compose to provide evidence for the system aswhole; difficulties of communication between stakeholders that use differentlanguages and practices in their disciplines; difficulties in dealing withadvanced concepts in AI, control and hardware design, software for criticalsystems; and others. As a consequence, autonomous systems are often developedusing advanced engineering techniques, but outdated approaches toverification.We propose a creative programme of work that will enable fundamental changesto the current state of the art and of practice. We will define amathematical framework that enables a common understanding of the diversepractices and concepts involved in verification of autonomy. Our frameworkwill provide the mathematical underpinning, required by any engineeringeffort, to accommodate the notations used by the various disciplines. Withthis common understanding, we will justify translations between languages,compositions of artefacts (engineering models, tests, simulations, and so on)defined in different languages, and system-level inferences fromverifications of components.With such a rich foundation and wealth of results, we willtransform the state of practice. Currently, developers build systems fromscratch, or reusing components without any evidence of theiroperational conditions. Resulting systems are deployed in constrainedconditions (reduced speed or contained environment, for example) or offeredfor deployment at the user's own risk. Instead, we envisage the futureavailability of a store of verified autonomous systems and components.In such a future, in the store, users will find not just systemimplementations, but also evidence of their operational conditions andexpected behaviour (engineering models, mathematical results, tests, and soon). When a developer checks in a product, the store will require all theseartefacts, described in well understood languages, and will automaticallyverify the evidence of trustworthiness. Developers will also be able to checkin components for other developers; equally, they will be accompanied byevidence required to permit confidence in their use.In this changed world, users will buy applications with clear guarantees oftheir operational requirements and profile. Users will also be able to askfor verification of adequacy for customised platforms and environment, forexample. Verification is no longer an issue.Working with the EPSRC TAS Hub and other nodes, and our extensive range ofacademic and industrial partners, we will collaborate to ensure that thenotations, verification techniques, and properties, that we consider,contribute to our common agenda to bring autonomy to our everyday lives.