Reinforcement Learning Control of a Biomechanical Model of the Upper Extremity

We address the question whether the assumptions of signal-dependent and constant motor noise in a full skeletal model of the human upper extremity, together with the objective of movement time minimization, can predict reaching movements. We learn a control policy using a motor babbling approach based on reinforcement learning, using aimed movements of the tip of the right index finger towards randomly placed 3D targets of varying size. The reward signal is the negative time to reach the target, implying movement time minimization. Our biomechanical model of the upper extremity uses the skeletal structure of the Upper Extremity Dynamic Model, including thorax, right shoulder, arm, and hand. The model has 7 actuated degrees of freedom, including shoulder rotation, elevation and elevation plane, elbow flexion, forearm rotation, and wrist flexion and deviation. To deal with the curse of dimensionality, we use a simplified second-order muscle model acting at each joint instead of individual muscles. We address the lack of gradient provided by the simple reward function through an adaptive learning curriculum. Our results demonstrate that the assumptions of signal-dependent and constant motor noise, together with the objective of movement time minimization, are sufficient for a state-of-the-art skeletal model of the human upper extremity to reproduce complex phenomena of human movement such as Fitts' Law and the 2/3 Power Law. This result supports the idea that the control of the complex human biomechanical system is plausible to be determined by a set of simple assumptions and can be easily learned.