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

CAREER: Robot Learning from Motor-Impaired Instructors and Task Partners PI, Northwestern University: Brenna D. Argall Overview Advances in robotics technologies are well-poised to make major contributions in the area of human as- sistance. Robots which attach to or support humans to provide physical assistance, and yet do not adapt to the varied and variable needs of their users, will however struggle to achieve widespread adoption and acceptance. Not only are the physical abilities of the user very non-static—and therefore also is their de- sired or needed amount of assistance—but how the user operates the robot too will change over time. The fact that there is always a human in the loop offers an opportunity: to learn from the human, transforming into a problem of robot learning from human teachers. Which raises a significant question: how will the machine learning algorithm behave when being instructed by teachers who not only are not machine learn- ing or robotics experts, but moreover have motor impairments that influence the learning signals which are provided? There has been limited study of robot learning from non-experts, who do not understand the details of how a given machine learning algorithm is working, and the domain of motor-impaired teachers is even more challenging. The proposed work contributes algorithmic approaches tailored specifically to the unique constraints of learning in this domain. We also contribute an evaluation of these algorithms in use by real end-users with motor impairments. Intellectual Merit The proposed work puts forth multiple hypotheses of ways in which constraints (like data sparsity, noise) can be advantageous for machine learning algorithms that intentionally exploit characteristics of the control and feedback signals provided by motor-impaired humans. Specifically, these advantages relate to recasting the problem into one that is more focused and tractable, how the task and motor data are encoded within the user’s control signals, the design of new formulations for robot behaviors and the interaction between the human teacher and robot learner. Rather than treat the constraints as limitations, we exploit them. Towards this end, we develop multiple novel machine learning algorithmic techniques, (1) that reason explicitly about the control interface to the robot and how it interacts with the full control space; (2) that derive, from noise in the human’s teleoperation commands, information about the human’s control patterns and the task requirements; and (3) which include the design of adaptation cues informed by reward- and example-based feedback from the motor-impaired teacher, in addition to autonomously computed metrics of team performance. Note that items (1) and (2) are of general utility to any domain that involves interaction between a human and a controlled system. Item (3) is of general utility to machine learning within real world domains, by advancing the ability to learn from imperfect or otherwise limited teachers. Moreover, we investigate a number of these hypotheses via subject studies with real motor-impaired end-users operating a robotic arm, both to explore this problem space and to assess the functionality and user preference and acceptance of the contributed algorithmic techniques. Broader Impacts The core pillars of the PI’s education plan include course development, mentorship, a student exchange and outreach. Specifically, the PI’s teaching mission is to design and teach courses that intersect robotics with machine learning and artificial intelligence,