Adaptation on Demand

Successful interaction with a changing environment requires learning from previous experience and applying such knowledge in adjusting motor commands in a predictive way. To investigate how control strategies are affected by internal models of the statistical properties of the environment, we tested subjects in reaching movements using a virtual reality system and perturbed the target in non-random ways during the movement. Perturbation time and dynamic environments were experimentally controlled to vary the necessity of using prediction of perturbations to make full corrections. We show that prediction of perturbations is used in on-line corrections only if it can lead to better endpoint positional accuracy and less energy consumptions. These results are well fit by an optimal feedback control model that enables us to handle non-Gaussian perturbation statistics and non-quadratic cost functions. In addition to demonstrating the flexible use of internal models in generating control strategies, we also show that different learning mechanisms are involved during motor learning, some of which are more complex than adjusting motor commands based on previous errors in a trial-by-trial manner.