By using an Extended Kalman Filter as our recursive estimator, we are able to incorporate a model of arm dynamics into the estimation process.
So far, we use a very simple (and inaccurate) model of the arm dynamics; we model the dynamics as a gaussian random walk in the joint velocities. The variance of the noise model is adjusted empiricially as a compromise between smoothness of tracking and sluggishness when changing the direction of motion.
In the near future, we will try to improve the predictive power of our estimator by making use of knowledge obtained from studies of human movement. For instance, it has been found that reaching movements are executed in a manner which minimizes joint torque. Also, despite the redundant number of degrees of freedom in the arm, there is a 'standard' arm pose to reach for a particular location.