How does the brain control movement? Experiments suggest that the (pre-)motor cortex behaves like an «engine» whose dynamics drive movement, and whose activity must first be initialised into movement-specific states (the «optimal subspace hypothesis», Shenoy et al, 2013). Both the computational and mechanistic underpinnings of this preparatory process remain poorly understood. Here, we propose a realistic circuit model for movement preparation and execution. We formalise movement preparation as an optimal control problem, under an internal forward model that predicts (future) patterns of muscle activity from momentary cortical preparatory states. We compute the preparatory input to cortex that drives fastest convergence to preparatory states predicted to yield the correct motor outputs. We also request that the control input keeps preparatory activity in an appropriate «output-null» subspace to prevent unwanted, premature motor outputs (Kaufman et al, 2014).

Critically, we show that optimal control inputs can be realised via feedback in realistic neural circuit architectures. Specifically, we model cortex as an inhibition-stabilised network, whose dynamics resembles those of monkey M1 during reaching (Hennequin et al, 2014). Optimal movement preparation is accomplished by a thalamo-cortical loop, gated by the basal ganglia. The loop is open by default, closed to drive movement preparation, and eventually re-opened to initiate movement. Importantly, we find that control loops can be flexibly combined to generate movements assembled from a few movement primitives.

The model produces naturalistic patterns of preparatory activity, including complex transients early during preparation, and displaying substantial variability in output-null dimensions. Consistent with data, across-trial variability is suppressed during preparation. Moreover, preparation may be as short as 200ms without loss of motor accuracy, also consistent with recent experimental observations. Our work brings together several threads of experimental research on both cortical and subcortical areas, and offers a new computational, normative perspective on the dynamics of motor circuits.