From Perception and Learning to Understanding and Action

长期以来，人工智能（AI）一直致力于创造能够观察世界并根据这些观察选择行动的智能代理。拟议的研究计划将通过探索与计算机视觉、具体智能、自然语言理解、深度学习、强化学习和机器人技术相结合的基础研究问题来实现。该研究计划将广泛使用模拟环境和人-代理交互来探索深度学习、深度强化学习和具有具身代理的问答。研究将探索基本问题以及现实设置，包括：自动驾驶汽车，与世界进行物理交互的机器人以及纯粹基于软件的信息代理。本研究将考察三种具有代表性的设置。(1)自动驾驶汽车：在这种情况下，上面讨论的许多抽象概念变得更加牢固地植根于具体任务。此外，与人工智能的许多子学科相关的关键问题，如计算机视觉、自动推理、规划和强化学习范式（RL），在这样的应用中可以变得更加清晰。(2)其他设置，例如与为工业和家庭设置创建更通用的机器人相关的设置，也可以作为接地应用。在这些设置中，代理对世界和世界中的物体如何与代理进行物理交互的理解也成为问题的重要元素。当人们开始设想人们可能与更通用和“具体化”的智能代理交互的方式时，自然语言作为促进人类与机器人交互的强大方式出现了。语言也可以作为教授代理新能力的机制。语言也可以与视觉演示相结合，以进一步增强具身代理的学习。(3)智能助手——比如那些通过智能手机或智能扬声器与用户互动的助手，可以被认为是存在于信息世界中的软件代理。实体代理人与信息代理人的结合可以将人工智能的好处推向加拿大人的日常世界。北美正在进行大规模的自动驾驶汽车测试，美国也开始运行没有安全驾驶员的汽车。这里提出的自动驾驶汽车导向研究探索了对加拿大和加拿大人在公共安全和经济繁荣方面至关重要的主题。此外，更多的通用机器人有可能改变制造、分销和交付服务，并为残疾人提供强大的辅助工具。

Artificial Intelligence (AI) has long concerned itself with the notion of creating intelligent agents capable of observing the world and selecting actions to take based on those observations. The proposed research program will be pursued through exploring fundamental research problems linked to the combination of computer vision, embodied intelligence, natural language understanding, deep learning, reinforcement learning and robotics. The research program will make extensive use of simulation environments and human--agent interaction to explore deep learning, deep reinforcement learning, and question answering with embodied agents. Research will explore fundamental questions as well as realistic settings including: autonomous vehicles, robots that physically interact with the world and purely software based information agents. This research will examine three representative settings. (1) Autonomous vehicles: in this setting, many of the abstract concepts discussed above become more firmly grounded in a concrete task. Furthermore, key problems associated with many sub-disciplines of AI such as computer vision, automated reasoning, planning and the paradigm of reinforcement learning (RL) can become clearer when grounded in such an application. (2) Other settings such as those associated with the creation of more general-purpose robotics for industrial and home settings can also serve as grounding applications. In these settings an agent's understanding of how the world and objects within the world physically interact with the agent also becomes an important element of the problem. As one begins to envision the way in which people might interact with more general purpose and "embodied" intelligent agents, natural language emerges as a powerful way to facilitate human--robot interactions. Language can also serve as a mechanism for teaching agents new capabilities. Language can also be coupled with visual demonstrations to further enhance the learning of an embodied agent. (3) Intelligent assistants - such as those that interact with the user through a smartphone or smart speaker, could be thought of as software agents that exist in a world of information. The combination of physically embodied agents with information agents could propel the benefits of AI into the everyday world of Canadians. The large scale testing of autonomous cars is underway in North America and cars without safety drivers are now operating in the United States. The autonomous car oriented research proposed here explores subjects of critical importance to Canada and Canadians in terms of both public safety and economic prosperity. Further, more general-purpose robotics has the potential to transform manufacturing, distribution and delivery services and well as provide powerful assistance tools for the disabled.