A realistic human kinematic model that satisfies anatomical constraints isessential for human-robot interaction, biomechanics and robot-assistedrehabilitation. Modeling realistic joint constraints, however, is challengingas human arm motion is constrained by joint limits, inter- and intra-jointdependencies, self-collisions, individual capabilities and muscular orneurological constraints which are difficult to represent. Hence, physiciansand researchers have relied on simple box-constraints, ignoring importantanatomical factors. In this paper, we propose a data-driven method to learnrealistic anatomically constrained upper-limb range of motion (RoM) boundariesfrom motion capture data. This is achieved by fitting a one-class supportvector machine to a dataset of upper-limb joint space exploration motions withan efficient hyper-parameter tuning scheme. Our approach outperforms similarworks focused on valid RoM learning. Further, we propose an impairment index(II) metric that offers a quantitative assessment of capability/impairment whencomparing healthy and impaired arms. We validate the metric on healthy subjectsphysically constrained to emulate hemiplegia and different disability levels asstroke patients.