CHS: Small: Watch One, Do One, Teach One: An Integrated Robot Architecture for Skill Transfer

在过去的几年里，机器人研究已经从仅仅关注构建完全自主和有能力的机器人转变为包括构建与人类合作伙伴协作的部分能力的机器人，从而使机器人能够做机器人最擅长的事情，而人类也可以做人类最擅长的事情。 这种转变是由旨在加强中小型制造的安全、交互式系统的复兴推动的，并且伴随着我们对机器人训练方式的改变。 机器人独立操作、从人们执行任务的被动观察中学习的学习机制正在被机器人通过与人类伙伴合作完成任务来学习的机制所取代。 该项目将寻求开发一种机器人架构，允许专家人类教练向机器人传授新技能，然后让机器人成为与人类伙伴并肩作战的熟练合作者，最后让机器人将学到的技能传授给人类新手学生。 为了实现这一目标，需要放弃流行但不透明的学习机制，转而采用新颖的表示方式，这种表示方式允许快速学习，同时在协作和教学过程中保持解释透明，同时认真考虑人类伙伴的心理状态（知识、目标和意图）。 这项工作的一个基本成果将是一个统一的表示，将现有的学习文献从演示到协作场景以及涉及机器人作为指导者的场景联系起来。因此，项目成果将对协作制造等应用领域产生广泛影响，同时也加强我们对教育和培训（特别是为研究生和本科生研究人员提供的研究服务）的大量投资，并将进一步丰富扩大计算参与的努力。这项工作将建立在三个子领域的研究基础上，并扩展最先进的技术，以解决每个领域的缺陷：1 - 机器人作为学生。 在“从演示中学习”工作的基础上，该团队将构建能够向人类学习任务模型的机器人。 然而，为了对其他领域有用，这些模型不能像许多当前的学习技术那样不透明。 相反，透明的模型将允许机器人提供并询问有关其性能的反馈，解释它所学到的知识，并主动提出加快​​学习速度的问题。2 - 机器人作为协作者。 相对较新的人机协作领域在人类和机器人合作伙伴之间同步任务执行方面遇到了困难。 通过链接到学习任务行为模型以及用户意图和理解模型，团队将构建能够熟练协商任务分配、适应用户偏好以及在出现错误或计划更改时恢复/更新内部表示的系统。3 - 机器人作为老师。 包括智能辅导系统在内的领域构建用户知识模型，通常使用贝叶斯知识追踪进行建模。 然而，这些模型只是简单地显示已知、未知或遗忘的知识，并且仅适用于事实知识。 通过与任务和意图的具体表示相联系，该团队将创建能够检测、扩展或修复学生执行现实世界任务的心理模型的机器人。三年的一系列里程碑将最终展示一个机器人，该机器人可以学习新任务，协作完成该任务，然后将该任务传授给其他人。该奖项反映了 NSF 的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

In the last several years, robotics research has transitioned from being concerned exclusively with building fully autonomous and capable robots to include building partially-capable robots that collaborate with human partners, allowing the robot to do what robots do best and the human to do what humans do best. This transition has been fueled by a renaissance of safe, interactive systems designed to enhance the efforts of small- and medium-scale manufacturing, and has been accompanied by a change in the way we think robots should be trained. Learning mechanisms in which the robot operates in isolation, learning from passive observation of people performing tasks, are being replaced by mechanisms where the robot learns through collaboration with a human partner as they accomplish tasks together. This project will seek to develop a robot architecture that allows for new skills to be taught to a robot by an expert human instructor, for the robot to then become a skilled collaborator that operates side-by-side with a human partner, and finally for the robot to teach that learned skill to a novice human student. To achieve this goal, popular but opaque learning mechanisms will need to be abandoned in favor of novel representations that allow for rapid learning while remaining transparent to explanation during collaboration and teaching, in conjunction with a serious consideration of the mental state (the knowledge, goals, and intentions) of the human partner. A fundamental outcome of this work will be a unified representation linking the existing literature in learning from demonstration to collaborative scenarios and scenarios involving the robot as an instructor. Thus, project outcomes will have broad impact in application domains such as collaborative manufacturing, while also enhancing our substantial investment in education and training (especially research offerings for graduate and undergraduate investigators), and will furthermore enrich the efforts to broaden participation in computing.This effort will build upon research in three subfields and extend the state-of-the-art to address deficiencies in each:1 - Robot as Student. Building on work from Learning from Demonstration, the team will construct robots that learn task models from humans. However, to be useful to the other thrust areas, these models must not be opaque as many current learning techniques are. Instead, a transparent model will allow the robot to provide and ask feedback about its performance, explain what it has learned, and to proactively ask questions that speed up learning.2 - Robot as Collaborator. The relatively new field of Human-Robot Collaboration struggles with synchronizing task execution between human and robot partners. By linking to models of learned task behavior and models of user intention and understanding, the team will construct systems that become proficient in negotiating task allocation, accommodating user preferences, and restoring/updating internal representations in case of errors or change of plans.3 - Robot as Teacher. Fields including Intelligent Tutoring Systems build models of user knowledge, typically modeled using Bayesian knowledge tracing. These models, however, simply show knowledge as known, unknown, or forgotten, and only for factual knowledge. By linking with concrete representations of task and intent, the team will create robots that can detect, extend, or repair the mental model of a student for real-world tasks.A set of milestones across three years will culminate in a demonstration of a robot that can learn a new task, collaborate on that task, and then teach that task to others.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.