Models and Plan Recognition Algorithms for Assistant Agents

该研究计划旨在开发新的模型和算法，以改进最先进的计划识别方法。计划识别是一种人工智能（AI）能力，它与规划相反。自动规划是生成计划（例如行动序列）的能力，它使智能系统能够自主地做出复杂的决策以实现其目标。计划识别是相反的。该系统通过传感器（如摄像头）观察代理（如人）的动作。从观察到的动作序列中，计划识别算法试图推断出代理的计划、目标甚至意图是什么。这是与其他代理（人类或非人类）交互的智能系统的基本功能。计划识别有许多应用领域：机器人、智能家居（帮助居民）、视频游戏（检测玩家意图）、网络安全（对抗黑客）、防御（预测敌人的威胁）等。本课题的研究重点是由认知辅助机器人驱动的。长期目标是设计能够帮助老年人在家中进行日常生活活动（ADL）的助理机器人，就像人类护理人员一样。为了实现这一目标，必须解决许多人工智能挑战。该项目的长期目标是为辅助机器人设计新的、更有效的计划识别技术和算法。这个项目有三个短期目标。第一个目标是提高计划识别方法的时间推理能力。我们建议通过考虑行动的持续时间和何时使用随机模型来解决许多歧义。第二个目标是将计划识别与另外两个功能，即运动规划和目标识别（图像分类）统一起来。我们不再单独考虑这些能力和被动的计划识别过程，而是将这些能力统一起来，以便更好地定位机器人以最大化其观察范围。最后，第三个目标旨在减少基于计划库的方法所需的知识量。我们将开发通过学习被观察主体的习惯来自动改进计划的方法。该计划具有很高的影响潜力。它的目标是由实际应用的具体需求所驱动的计划识别方面的基本进步。在实践中，这将提高辅助机器人的能力。加拿大面临着严重的人口老龄化问题，这给包括卫生在内的公共服务带来了压力。这些服务的成本正在上升，我们面临着严重的劳动力短缺。机器人助手等技术解决方案将有助于克服人口方面的挑战。帮助老年人在家中延长寿命也能提高他们的生活质量。

This research program aims to develop new models and algorithms to improve state-of-the-art methods in plan recognition. Plan recognition is an Artificial Intelligence (AI) capability, which is the is the opposite of planning. Automated planning is the capability of generating plans (e.g. sequences of action), which enables intelligent systems to make complex decisions to achieve their goals autonomously. Plan recognition is the opposite. The system observes the actions of an agent (e.h., a person), using sensors (e.g., a camera). From the observed sequence of actions, a plan recognition algorithm attempts to deduce what the agent's plan, goal, or even intentions are. This is an essential capability for intelligent systems that interact with other agents (humans or not). Plan recognition has many fields of application : robotics, smart homes (helping residents), video games (detecting player intentions), cybersecurity (countering hackers), defence (anticipating threats from the enemy), etc. The research focus of this program is specially motivated by cognitive assistance robots. The long-term goal is to design assistant robots that can help elderly people in their activities of daily living (ADL) at home, much like a human caregiver. To achieve this goal, many AI challenges must be addressed. The long-term objective of the program is to design new and more efficient techniques and algorithms of plan recognition for assistant robots. This program has three short-term objectives. The first objective is to improve the temporal reasoning capabilities of plan recognition methods. We propose to resolve many ambiguities by considering the duration of the actions and when they are executed using stochastic models. The second objective aims to unify plan recognition with two other capabilities, namely motion planning and object recognition (image classification). Instead of considering these capabilities independently and a passive plan recognition process, we will unify these capabilities to better position the robot to maximize its observations. Finally, the third objective aims to reduce the amount of required knowledge of plan libraries-based methods. We will develop methods to automatically refine plans by learning the habits of the observed agents. This program has a high impact potential. It targets fundamental advances in plan recognition motivated by concrete needs of real applications. In practice, this will improve the abilities of assistant robots. Canada faces a major aging of its population which puts pressure on public services, including health. The cost of these services is increasing and we face a serious labour shortage. Technological solutions, such as robot assistants, will help to overcome demographic challenges. Helping older people live longer at home also improves their quality of life.