A safety-assured architecture for AI-enabled autonomous vehicles

人工智能（AI）使新型自主移动机器人成为可能，这些机器人即将改变社会的多个领域。其中包括自动驾驶汽车（AVs）、人行道送货机器人、自动农业机械、清洁机器人、自动火车和无人机。如今，许多机器人从根本上依赖于深度学习来完成物体识别等感知任务，但深度学习也开始渗透到预测、规划和控制任务中。这些进步极大地提高了这些系统的适应性及其处理复杂情况和相互作用的能力。这些系统有望在开放环境中与人类和其他机器人进行交互，这些环境会受到意想不到的障碍、变化的天气条件和不断发展的基础设施的影响。不幸的是，依赖深度学习的系统对安全保障提出了巨大挑战，这是其广泛部署的主要障碍。深度神经网络（dnn）缺乏人类的可解释性，并且可能以不可预测的方式在新的输入上失败。此外，他们的目标任务，如识别行人，逃避了完整的规范。这些特性使得传统的安全保证方法难以或不可能应用于这些系统。该发现基金将资助一项前沿研究项目，为自动驾驶汽车创建下一代系统架构，这些系统架构将利用现代人工智能技术，但本质上是可靠的。预期结果包括：(1)感知和规划架构，该架构受到人类认知双过程理论的启发，并以可解释的符号形式将dnn的高性能决策与安全相关推理集成在一起；(2)采用该体系结构保证机器人安全的方法。该项目将使用由PI团队开发的UW Moose自动驾驶汽车，在公共道路上演示和评估这项新技术。自动驾驶汽车和其他移动机器人将在未来几十年对我们的社会和经济产生重大影响。预计自动驾驶汽车可以将事故率降低80%。在加拿大，这将每年挽救1600人的生命，减少超过550亿美元的医疗成本和生产力损失。拟议的研究计划针对目前更广泛采用此类系统的主要障碍，即确保其安全性。该项目将产生人工智能和机器人领域的前沿研究成果，共享数据集和基准，以刺激科技进步，创造商业化技术，并在未来5年内培养24名人工智能和机器人领域的HQP。在自动驾驶汽车上的实践经验将使这些HQP在就业市场上具有显著的竞争优势。汽车行业是加拿大制造业GDP的最大贡献者。加拿大在自动驾驶技术方面的创新和HQP培训将有助于保持加拿大在汽车行业和人工智能领域的世界领先地位。

Artificial intelligence (AI) has enabled new kinds of autonomous mobile robots that are on the verge of transforming multiple sectors of society. These include autonomous vehicles (AVs), sidewalk delivery robots, autonomous agricultural machines, cleaning robots, autonomous trains, and drones. Today, many of these robots fundamentally rely on deep learning for perception tasks such as object recognition, but deep learning starts to also penetrate prediction, planning, and control tasks. These advances greatly improve the adaptability of these systems and their ability to handle complex situations and interactions. These systems are expected to interact with humans and other robots in open environments that are subject to unexpected obstacles, varying weather conditions, and evolving infrastructures. Unfortunately, systems that rely on deep learning pose a great challenge to safety assurance, which is a major obstacle to their wide deployment. Deep neural networks (DNNs) lack human interpretability and may fail in unpredictable ways on new inputs. Further, the tasks they target, such as recognizing a pedestrian, evade complete specifications. These properties make the application of traditional safety assurance methods to these systems difficult or impossible. This Discovery grant will fund a cutting-edge research program to create the next generation of system architectures for AVs that take advantage of modern AI-technologies but are inherently assurable. The expected results include (1) an architecture for perception and planning that is inspired by the dual-process theory of human cognition and integrates the high-performance decision-making of DNNs with safety-relevant reasoning in interpretable, symbolic form; and (2) the methods to assure the safety of robots using such an architecture. The program will use the UW Moose, an AV developed by the PI's team, to demonstrate and evaluate the new technology on public roads. AVs and other mobile robots will have a major impact on our society and economy over the coming decades. It is expected that AVs can reduce accident rates by as much as 80%. In Canada, this would save 1,600 lives a year and over $55 billion in healthcare costs and lost productivity. The proposed research program targets current major roadblock to a wider adoption of such systems, which is assuring their safety. The program will generate cutting-edge research results in AI and robotics, share datasets and benchmarks to stimulate scientific and technological progress, create commercializable technology, and train 24 HQP in AI and robotics over the next five years. Hands-on experience on an AV will give these HQP a significant competitive advantage on the job market. The automotive sector is the single biggest contributor to Canada's manufacturing GDP. Canadian innovation and HQP training in AV technology will contribute to maintaining Canada's role as a world leader in the automotive sector and AI.