CAREER: Robot Learning from Motor-Impaired Instructors and Task Partners

辅助机器--如电动轮椅、机械臂、外骨骼、假肢--对于使有严重运动障碍的人能够独立至关重要。然而，存在着一个矛盾的现象，一个人的损伤越严重，他们就越不能操作这些可能改善他们生活质量的机器。在这里，机器人技术有可能改变人类健康和康复领域：通过将机器变成机器人，可以自主操作并分担控制负担。这些机器人必须适应人类用户的独特偏好和能力，而这两者如何随着时间的推移而变化，对于实现广泛的采用和接受至关重要，特别是如果连接到人体上提供物理帮助。对于机器人从非专家那里学习的研究有限，而运动障碍教师的领域更具挑战性：他们的控制信号是嘈杂的（由于运动信号中的伪影）和稀疏的（如果提供运动命令是更多的努力），并通过界面过滤。而不是把这些约束作为限制，提出的工作假设，这样的约束变得有利于机器学习算法，利用独特的特征（如问题空间稀疏）的控制和反馈信号从运动受损的人。这项工作开发了多种新的机器学习算法技术，（1）明确地推理控制界面以及它如何与整个机器人控制空间交互;（2）从人类遥控操作命令的变化中获得有关人类控制模式和任务要求的信息;（3）包括由运动障碍教师提供的基于奖励和样例的反馈信息设计适应线索。拟议的工作还与电机受损的最终用户操作多个机器人平台进行了主题研究，以探索这个问题空间，并评估所贡献的算法技术的功能和用户接受度。

Assistive machines - like power wheelchairs, robotic arms, exoskeletons, prostheses - are vital for enabling independence for people with severe motor impairments. However, there exists a paradox, where often the more severe a person's impairment, the less able they are to operate these very machines which might improve their quality of life. Here robotics technologies have the potential to transform the field of human health and rehabilitation: by turning the machine into a robot, that can operate itself autonomously and share the control burden. It will be crucial that these robots adapt to the human user's unique preferences and abilities, and how both change over time is crucial for achieving widespread adoption and acceptance, and especially if attached to a human's body to provide physical assistance. There has been limited study of robot learning from non-experts, and the domain of motor-impaired teachers is even more challenging: their control signals are noisy (due to artifacts in the motor signal) and sparse (if providing motor commands is more effort), and filtered through an interface. Rather than treat these constraints as limitations, the proposed work hypothesizes that such constraints become advantageous for machine learning algorithms that exploit unique characteristics (like problem-space sparsity) of control and feedback signals from motor-impaired humans. The work develops multiple novel machine learning algorithmic techniques, (1) that reason explicitly about the control interface and how it interacts with the full robot control space; (2) that derive information about the human's control patterns and task requirements, from variability in the human's teleoperation commands; and (3) which include the design of adaptation cues informed by reward- and example-based feedback from motor-impaired teachers. The proposed work also performs subject studies with motor-impaired end-users operating multiple robotic platforms, both to explore this problem space and assess the functionality and user acceptance of the contributed algorithmic techniques.