CPS: Medium: GOALI: Enabling Safe Innovation for Autonomy: Making Publish/Subscribe Really Real-Time

在当今的汽车行业，各公司正在激烈竞争，希望在其产品线中引入更复杂的自动驾驶功能。这场竞争有望达到高潮，那就是在大众市场规模上实现完全自治。这里的赌注很高：最先到达那里的公司(和国家)将处于主导地位，影响与自治相关的能力在未来几十年的演变。这种高风险的竞争导致了在感知和决策能力等自动驾驶关键技术方面迅速创新的巨大压力。这种压力导致了系统设计的“黑箱”方法，将原本用于其他环境的现成软件和硬件组件重新用于实现自动驾驶功能。ROS(机器人操作系统)是使用最广泛的重新调整用途的黑盒组件之一。ROS使单独开发的软件程序能够实现不同的功能(例如，基于摄像头的感知、正确的车道跟踪等)。组合在一起形成一个提供更广泛功能的系统(例如，自动驾驶的汽车)。不幸的是，顾名思义，ROS最初的设计和实现是为了支持机器人应用程序的开发，这些应用程序的要求与自动驾驶汽车非常不同。因此，ROS缺乏确保安全的汽车系统设计所需的功能。这里的一个关键问题是缺乏对确保实时安全的支持，即某些功能(例如，刹车)是“按时”(例如，在遇到障碍物之前)执行的。该项目旨在开发一种将实时安全作为首要问题的ROS的替代方案。尽管其名称不同，但ROS实际上不是一个操作系统(OS)，而是一组用户级中间件库，有助于构建典型的机器人应用程序的过程图。这些库通过图形节点之间消息通信的发布/订阅(发布/订阅)概念支持模块化系统开发，从而允许不同的软件包松散耦合。这种松散耦合实现了软件重用，这是ROS在实现快速创新方面取得成功的关键。ROS的成功令人信服地证明了酒吧/订阅在推动自主创新方面的重要性。但是，发布/订阅必须是安全的才能申请。该项目针对的正是这个问题，特别是在自动驾驶汽车中使用的多核+加速器平台的背景下。在这样的平台中，仅使用CPU的多核计算机增加了图形处理单元(GPU)等协处理器，这些协处理器可以加速基于AI的软件中常见的某些数学计算，以实现自主性。该项目的具体目标是生产一种PUB/SUB替代ROS，便于进行实时安全认证。主要研究任务包括解决操作系统和中间件级别的基本资源分配问题，在实时发布/订阅图中生成用于验证响应时间界限的分析，生成参考发布/订阅中间件实现，并在实验上将该实现与ROS进行比较。虽然不断演变的ROS本身超出了本项目的范围，但该项目将揭示与此类演变相关的基本权衡。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

In the automotive industry today, companies are fiercely competing to field ever more sophisticated autonomous features in their product lines. The hoped-for culmination of this competition is full autonomy at mass-market scales. The stakes here are high: the companies (and countries) that get there first will be in a commanding position to influence how autonomy-related capabilities evolve for decades to come. This high-stakes competition has resulted in significant pressure to innovate quickly with respect to key technologies for autonomous driving, such as perception and decision-making capabilities. This pressure has led to a “black-box” approach to system design, with off-the-shelf software and hardware components, originally intended for other contexts, repurposed to implement autonomous-driving functions. One of the most widely used repurposed black-box components is ROS (the Robot Operating System). ROS enables separately developed software programs that implement different functions (e.g., camera-based perception, correct lane following, etc.) to be combined to form a system that provides broader capabilities (e.g., a car that drives itself). Unfortunately, as its name suggests, ROS was originally designed and implemented to support the development of robotics applications, which have very different requirements from autonomous vehicles. As a result, ROS lacks features needed to ensure safe automotive system designs. A key issue here is a lack of support for ensuring real-time safety, i.e., that certain functions (e.g., braking) are performed “on time” (e.g., before an obstacle is hit). This project is directed at producing an alternative to ROS that takes real-time safety as a first-class concern.Despite its name, ROS is really not an operating system (OS) but rather a set of user-level middleware libraries that facilitate constructing processing graphs typical of robotics applications. These libraries support modular system development via a publish/subscribe (pub/sub) notion of message communication between graph nodes that allows different software packages to be loosely coupled. This loose coupling enables software reuse, which has been a key to ROS’s success in enabling rapid innovation. ROS’s success convincingly demonstrates the importance of pub/sub in fueling innovation in autonomy. However, pub/sub must be safe to apply. This project is directed at this very issue, specifically in the context of multicore+acclerator platforms as used in autonomous vehicles. In such a platform, a CPU-only multicore computer is augmented with co-processors like graphics processing units (GPUs) that can speed up certain mathematical computations that commonly occur in AI-based software for autonomy. The specific aim of this project is to produce a pub/sub alternative to ROS that facilities real-time safety certification. Key research tasks include resolving fundamental resource-allocation concerns at the OS and middleware levels, producing analysis for validating response-time bounds in real-time pub/sub graphs, producing a reference pub/sub middleware implementation, and experimentally comparing this implementation to ROS. While evolving ROS itself is beyond the scope of this project, this project will expose fundamental tradeoffs of relevance to such an evolution.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.