An Upper Limb Kinematic Graphical Model for the Prediction of Anthropomorphic Arm Trajectories

This study approaches the use of Bayesian networks to model the human arm movement in an anthropomorphic manner for the control of an upper limb assistive robot. The model receives as input a desired wrist position and outputs three angles, the swivel angle (i.e. the angle that represents the rotation of the plane formed by the upper and lower arm around the axis that passes through the shoulder and wrist) and two angles corresponding to two degrees of freedom of the sternoclavicular joint (elevation/depression and protraction/retraction). These angles, together with the wrist position, fully describe the position of the shoulder and the elbow. A set of recording sessions was conducted to acquire human motion data to train the model for four different activities of daily living. Performance was measured by the elbow and shoulder joints' end-point errors and Pearson's r. The model was able to predict accurately elbow movement (mean error $\pmb{0.021}\pm\ \pmb{0.020}\mathbf{m}$, Pearson's $r$ 0.86-0.99) and shoulder movement (mean error $\pmb{0.014}\pm \pmb{0.011}\mathbf{m}$, Pearson's $r$ 0.52-0.99) for wrist trajectories that fall in the set of training data. It was also able to create new motions that were not in the set of training data, with a better accuracy for the elbow joint (mean error $\pmb{0.042}\pm \pmb{0.025}\mathbf{m}$, Pearson's $r$ 0.59-0.99) and an average accuracy for the shoulder joint (mean error $\pmb{ 0.026}\pm \pmb{0.012}\mathbf{m}$, Pearson's r −0.12-0.99). The proposed model presents a novel method to solve the inverse kinematics problem in the scope of the human upper limb. It can also create movement out of its training data, although not highly correlated with the trajectory performed by a human.