 This paper examines the potential of robotic and virtual reality-based rehabilitation and automated assessments using data-driven learning. The authors found that deep learning with spatiotemporal skeleton data and dynamic attention outperformed other approaches, achieving an RMSE, root mean square error, of 0.55. Automating rehabilitation can speed up objective assessment and increase accessibility for more people. This article was authored by Sejuti Ramon, Sujan Sarkar, A.K.M. Natimal Hack and others.