Feedback control of random networks as a model of flexible motor cortical dynamics across tasks.

Modern large-scale recordings show that motor cortex activity during reaching evolves along low-dimensional trajectories, often interpreted as neural bases for motor command generation. However, the origin of these patterns and how they flexibly reorganize across tasks remains unclear. Here, we show that hallmark features of motor cortical activity, including the encoding of several movement parameters during preparation and rotational dynamics during execution, emerge naturally in a linear model composed of a random network coupled to a biomechanical system. Under optimal control, the model achieves multiple behaviors through task-dependent mapping of sensory feedback onto the network. The corresponding feedback rule enables a rich motor repertoire while producing network activity consistent with motor cortical recordings. In this model, the relationship between feedback signals from the biomechanical plant and the low-dimensional network dynamics can be established analytically, which offers an interpretable framework to link neural dynamics with sensorimotor control across tasks.