A Socially Assistive Robot Exercise Coach for the Elderly

Fasola & Matarić. A SAR Exercise Coach for the Elderlysegmented image. The motion frame, once thresholded, serves as the segmentation inputin the first step of the algorithm, and the user arm angles can be determined in the samemanner as described. Using motion instead of color segmentation, along with Kalmanfiltering of the positions of the hands and elbows, provides a more general recognitionapproach for domain-independent use, even though the overall algorithm remains thesame.The development of the SAR exercise system and visual recognition procedurepredated the availability of the Microsoft Kinect. Our future implementations of thesystem will utilize Kinect-type 3D vision technology and thus do away with the curtainand the planar limits of the motions. Nevertheless, the 2D nature of the exercises was notnoted as an issue by any of the participants in our user studies. Furthermore, the real-timevision algorithm described is general and can be applied to a variety of domains,including those wherein use of the Kinect is not appropriate or feasible (e.g., outdoors insunlight).3.6 BehaviorsThe behavior module represents the main loop of the SAR exercise coach system. Itcommunicates with all of the other system modules and manages the flow of theinteraction, including transitioning between the different exercise games, managing thesession and game clocks, saving user performance and session statistics to the database,and handling user-requested breaks, among other session-related tasks.3.6.1 Interaction Flow and Behavior ManagementEach of the four exercise games is managed by its own game behavior. In accordancewith our principle of encouraging users to be intrinsically motivated to engage in the task,the behavior module transitions between game types every 1–2.5 minutes in order todiminish the possibility of user boredom, which would negatively affect motivation level.The behavior module chooses which game to transition to based on a predefined sessionschedule, which specifies the order of the games to play and their respective durations.The session schedule is defined a priori by the experimenter such that each session’sorder of games is unique. This requirement is enforced to reduce the predictability of theexercise regimen for the user, which might negatively affect his or her perception of thesystem (Bickmore, Schulman, & Yin, 2010). The session schedule, however, thoughunique for each individual session, follows approximately the same structure throughoutall of the sessions. Specifically, the Workout, Sequence, and Memory games are eachallotted approximately 25% of the total session time (75% for all three games), with theImitation game being allotted 15% of the total session time and the remaining 10% beingallocated for user breaks, transition behaviors, and the introduction and conclusion of thesession. The Imitation game receives a lower allotment compared to the other threegames due to the fact that the user is in control of the exercise. Although the Imitationgame promotes user autonomy and users often find the game stimulating, in pilot studieswe’ve observed users may get bored if the game goes on for too long. Thus the Imitationgame is usually only played for 1–1.5 minutes at a time during the course of the exercisesession. If a session schedule is not set, or is incomplete, the robot chooses a game atrandom to play at regular game change intervals. The exact durations for each game weredetermined empirically by taking into account multiple factors, including the totalduration of the session (20 minutes), the allotted game time ratio, the time required for15