A Collaborative and Adaptive Feedback System for Physical Exercises

Maintaining motivation to meet physical exercise goals is a big challenge in virtual/home-based exercise guidance systems. Lack of motivation, long-maintained bad daily routines, and fear of injury are some of the reasons that cause this hesitation. This paper proposes a reinforcement learning-based virtual exercise assistant capable of providing encouragement and customized feedback on body movement form over time. Repeated arm curls were observed and tracked using single and dual-camera systems using the Posenet pose estimation library. To accumulate enough experience across individuals, the reinforcement learning model was collaboratively trained by subjects. The proposed system is tested on 36 subjects. Behavioral changes are apparent in 31 of the 36 subjects, with 31 subjects reducing movement errors over time and 15 subjects completely eliminating the errors. The system was analyzed for which types of feedback provided the highest expected value, and feedback directly related to the previous mistake provided the highest valued feedback (
$p < 0.0133$
). The result showed that the Reinforcement Learning system provides meaningful feedback and positively impacts behavior progress.