The purpose of this study was to examine the neural activity underlying an implicit motor learning task. In particular, our goals were to determine whether initial phases of procedural learning of a motor task involve areas of the brain distinct from those involved in later phases of learning the task, and what changes in neural activity coincide with performance improvement. We describe a novel integration of robotic technology with functional brain imaging and its use in this study of implicit motor learning. A portable robotic device was used to generate forces that disturbed the subjects' arm movements, thereby generating a "virtual mechanical environment" that the subjects learned to manipulate. Positron emission tomography (PET) was used to measure indices of neural activity underlying learning of the motor task. Eight health, right-handed male subjects participated in the study. Results support the hypothesis that different stages of implicit learning (early and late implicit learning) occur in an orderly fashion, and that distinct neural structures may be involved in these different stages. In particular, neuroimaging results indicate that the cortico-striatal loop may play a significant role during early learning, and that the cortico-cerebellar loop may play a significant role during late learning.