Observing and Clustering Coaching Behaviours to Inform the Design of a Personalised Robotic Coach

Abstract

Adherence to repetitive rehabilitation exercises is important in motor recovery after stroke. Similarly, repetitive solo practice exercises can improve the skill level of sports players. In both of these scenarios, regular human coaching has benefits, but in practice, the required training is often carried out alone, resulting in lowered adherence. This work presents a mixed methodology approach, novel in the context of designing for HRI, towards informing the design of a personalised robotic coach for stroke rehabilitation and squash. Using observations of human-human interactions, we first obtained action sequences of behaviours exhibited by coaches and physiotherapists. We then clustered these action sequences into behaviour graphs, with each graph representing a coaching policy usable for robotic control. Next we obtained coaches' and physiotherapists' reflections on the graphs' applicability to the real world. Finally, we provide an explanation of how the policies visualised in these graphs could be used for robotic control.