Flanagan, R. R., Vetter, P., Johansson, R. S., & Wolpert, D. M. (2003). Prediction precedes control in motor learning. Current Biology, 13(2), 146–150. doi:10.1016/S0960-9822(03)00007-1

Summary

In this paper, Flanagan et al. describe an experiment in which participants must grip an object with their index finger and thumb, and move it in a straight line to a target. The dynamics of the object were modified so that when they moved it in the horizontal plane, a proportional vertical force was applied to the object. Thus, to learn to move it in a straight line, participants had to adapt to the vertical force.

Flanagan et al. found that participants took a long time (about 70 trials) before they were able to fully adjust their trajectories to be straight. However, they took much less time (about 10 trials) to adjust the force with which they gripped the object to match that of the corresponding load force. These results suggest that there are two internal models (one for the grip force, and one for the arm trajectory) that are being learned at separate rates. Specifically, Flanagan et al. suggest that in the first case, it is a forward kinematic model that is being learned, while in the second case, it is a inverse dynamics model that is being learned. This is consistent with the demands of the task: the novel dynamics of the object require learning a new mapping from desired trajectory to motor commands (the inverse model), but they do not require learning a new mapping for controlling the load force. Rather, the motor system needs only to predict the load force so that it can appropriately adjust for it.

Takeaways

This paper basically answers the question I ended with in Kawato’s review: learning operates independently in the forward and inverse models. Flanagan et al. suggest that, computationally, this may be able to be explained by something like distal teacher learning.