Kalidindi, H. T.; Crevecoeur, F.

Modern large-scale recordings have revealed that motor cortex activity during reaching follows low-dimensional dynamics, thought to reflect sensorimotor computations. However, the origin of these patterns, and how they flexibly reorganize across different tasks remain unclear. Here we demonstrate that the key features of neural activity naturally emerge in a linear model combining a random network with a biomechanical system. Remarkably, this model shows how a fixed network can flexibly control different behaviors by optimally mapping sensory and internal state onto task-specific network inputs. Finally, analytical decomposition of the controller reveals that low-dimensional network dynamics directly follows from propagation of low-dimensional feedback signals from the biomechanical plant through the network. These results provide a biologically plausible mechanism for flexible motor control in the nervous system which directly links neural population dynamics to behavior.