This paper develops a body-machine interface for the control of a powered wheelchair using upper-body motion. Our goal was to infer a cursor's kinematics from the signals recorded from 4 Inertial Measurement Units placed on a subject's shoulders. We specified a Kalman filter measurement model that assumes the Euler angles, angular velocities, and linear accelerations of the shoulders are a stochastic linear function of the position, velocity, and acceleration of the virtual cursor. This model learned a system that encodes cursor movement along with training data. Experimental results show that taking advantage of the redundancy of the signal improves performance during a center-out reaching task. The resulting algorithm provides a platform for people with high-tetraplegia to communicate their intended motor actions with the environment using specialized assistive devices.