Practitioner Evaluation Of Machine Learning Therapeutic Exercise Robot

Novel technological approaches to home exercise programs are needed to offset the supply-demand gap for older adults, people with disability, rural residents and others who have limited access to physical and occupational therapy services. PURPOSE: The purpose of this study was to develop machine learning algorithms that allow a robot to learn upper-extremity exercises from therapist demonstrations, recreate them in a lifelike manner, and for practitioners to evaluate the accuracy and utility of the robot to lead therapeutic exercise training. METHODS: Video motion capture was used to obtain 3D upper-extremity kinematics from a therapist demonstrating a battery of shoulder exercises. A dynamical system based learning from demonstration algorithm was used to teach an ABB-YuMi robot how to perform shoulder exercises using the 3D kinematic data of the therapist. Through an online survey, 110 licensed therapists from across the U.S. were shown videos of the therapist’s demonstration and robot’s performance and asked to rate congruency of joint angles, velocity and fluidity of movements and qualitatively evaluate the potential for robot-led therapy in home settings. RESULTS: Therapists rated the overall accuracy as 88% [95% CI = 86, 90] for forward raise, 87% [85, 89] for lateral raise, 84% [82, 86] for shoulder press, and 84% [82,86] for internal and external rotation with similar levels of agreement for joint angle, velocity and fluidity of motion. Qualitative thematic analysis revealed that while therapists thought the movement agreement was good, they noted specific variations in velocity, fluidity, joint angle, range of motion, and transitions in the robot. Therapists reported the robot would be beneficial for home exercise programs because of its ability to provide visual demonstrations, particularly for rural areas, but expressed concerns about cost, technological limitations, practicality, anthropometric differences, inability to give real-time feedback, and loss of human connection. CONCLUSION: A learning from demonstration enabled therapeutic exercise robot has good potential as an in-home adjunct therapy to supplement clinical exercise programs, but further refinement of robot anthropomorphism, movement quality, and the addition of real-time feedback are needed for widespread adoption. Supported by: NSF Grant IIS 1830597 - National Robotics Initiative 2.0 - Ubiquitous Collaborative Robots