NRI: Mutually Assistive Robotics

这个项目将推动为残疾用户提供帮助的机器人如何与这些用户互动并向他们学习的最先进水平。目前的大多数工作对残疾采取了基于赤字的方法，在这种方法中，机器人被认为是互动中更有能力的合作伙伴，残疾用户为一个主要帮助他们完成日常生活任务的系统提供高水平的目标。这种方法在两个重要方面存在缺陷：第一，它去除了用户的代理和控制，直接抵消了辅助技术的心理好处，并有可能用失去对机器人的独立性来取代对照顾者的独立性的丧失；第二，它未能解决提高生活质量的任务，如艺术表达或美容，在这些任务中，具体的动作序列、执行这些动作的方式以及对这些动作的控制是目标本身。这项研究对辅助机器人采取了基于优势的方法，开发了允许机器人和用户自由地相互协助完成任务的新方法，并在提高人们生活质量的活动中对这些方法进行了评估，在这些活动中，用户对完成任务的目标和方式的自主性和控制力都很重要。项目成果将包括机器人学习的新方法，使残疾人能够协作设计、控制和影响机器人行为，同时从事令人愉快的爱好，控制自己的外表，并总体上参与与世界的创造性互动。这些方法将有助于确保下一代辅助机器人支持残疾人的生活质量和快乐的自我表达，以及他们的日常家务。这应该会显著改善所有年龄段的大量肢体残疾美国人的生活。人-机器人交互的算法将产生更广泛的影响，这些算法不仅显著推进该领域，而且为交互强化学习、机器人学习和智能辅助技术的未来工作提供信息。利用团队在辅助技术、人-机器人交互、增强现实和人-机器人交互方面的专业知识，将通过三项技术创新实现项目目标。首先，在从直接控制到语言的多个抽象层次上，通过机器人对人和人对机器人的互助来增强初始模型学习的算法。第二，为用户提供可用的机器人心理模型的新方法，如通过增强现实来选择和显示信息，使用户能够理解机器人的感知和决策，并提高他们影响机器人行为的能力。最后，新的交互式学习算法使用户能够利用初始学习后的反馈，并确保用户可以影响任务执行的方式和任务目标。这些算法将统一在辅助机器人的三层体系结构中，明确支持从机器人到人类以及从人类到机器人的每个级别的辅助。最低层是数据驱动的从感觉状态到运动的映射。在中间级别，这些动作被命名为原子动作，例如到达、倾倒或抓取，并基于诸如目标对象或运动特征等参数进行分组。在最高级别，操作用前置条件和后置条件象征性地表示，并合并到多步骤计划中，这些计划实现用户指定的目标，同时由用户在线修改。除了在所有各级支持机器人对人类和人类对机器人的援助外，这一架构还将允许各级之间的信息流动，特别是在机器人对人类的援助中。这项工作将在专家用户-残疾人合作者的帮助下进行验证，以及在验证基础技术开发的更大规模研究中进行验证。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。

This project will advance the state-of-the-art for how robots that render assistance to users with disabilities interact with and learn from those users. Most current work takes a deficit-based approach to disability, in which the robot is assumed to be the more competent partner in the interaction and the user with disabilities provides high-level goals to a system that primarily helps them with mundane tasks of daily living. This approach is deficient in two significant ways: first, it removes agency and control from users, directly counteracting the psychological benefits of assistive technology and potentially replacing a loss of independence to caregivers with a loss of independence to robots; and second, it fails to address tasks that improve quality of life, such as artistic expression or grooming, where the specific sequence of actions, the manner in which those actions are carried out, and control over those actions is the goal itself. This research takes a strengths-based approach to assistive robotics, developing new methods that allow the robot and user to freely assist each other to complete tasks, and evaluating those methods in activities that improve people's quality of life and where users' autonomy and control over both the goal and manner of completing a task are important. Project outcomes will include new methods for robot learning that empower people with disabilities to collaboratively design, control, and influence robot behavior while engaging in pleasurable hobbies, controlling their own appearance, and generally engaging in creative interaction with the world. These methods will help to ensure that the next generation of assistive robotics support the quality of life and joyous self-expression for people with disabilities, as well as their daily chores This should significantly improve the lives of the substantial number of Americans of all ages who live with physical disabilities. Additional broad impact will derive from algorithms for human-robot interaction that significantly advance not only that field but also inform future work in interactive reinforcement learning, learning for robotics, and intelligent assistive technologies.Leveraging the team's expertise in assistive technology, human-robot interaction, augmented reality, and human-robot interaction, project goals will be achieved through three technological innovations. First, algorithms to enhance initial model learning with mutual assistance from robot to human and human to robot at multiple levels of abstraction, from direct control to language. Second, new methods for giving users usable mental models of the robot, such as selecting and displaying information through augmented reality to empower users to understand robot perception and decision-making and improve their ability to influence robot behavior. Finally, new interactive learning algorithms that enable users to exploit feedback after initial learning and ensure that users can influence the manner in which tasks are conducted as well as task goals. These algorithms will be united in a three-layer architecture for assistive robotics that explicitly supports assistance from both robot to human and human to robot at each level. At the lowest level is a data-driven mapping from sensory state to movement. At the middle level, those motions are named as atomic actions such as reaching, pouring, or grasping, and grouped based on parameters such as target objects or features of the motion. At the highest level, actions are represented symbolically with pre- and post-conditions and combined into multi-step plans that achieve user-specified goals while being modified online by the user. In addition to supporting both robot-to-human and human-to-robot assistance at all levels, this architecture will also allow for the flow of information between levels, especially in robot-to-human assistance. The work will be validated with the help of expert user-collaborators with disabilities, as well as in larger-scale studies that validate foundational technological developments.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.