How do humans learn to control movements in uncertain environments? This research presents an extension of the modular selection and identification for control (MOSAIC) model, a modular architecture for motor learning and control based on multiple pairs of forward (predictor) and inverse (controller) models. Learning in the architecture was implemented using both gradient-descent and the expectation-maximization (EM) algorithm. Simulations of object manipulation prove the architecture can learn to manipulate multiple objects and switch between them appropriately. Moreover, the model shows generalization to novel objects. Finally, when dynamics is associated with a particular object shape, the model selects the appropriate controller before movement execution. This research provides insights into the neural mechanisms underlying sensorimotor control and has implications for robotics and rehabilitation. Activation of modules is followed by on-line correction.
This paper, published in Neural Computation, fits squarely within the journal’s focus on computational neuroscience and machine learning. The extension and evaluation of the MOSAIC model for sensorimotor learning aligns with the journal’s emphasis on understanding neural mechanisms through computational modeling, making it highly relevant to the readership.