Sensorimotor control occurs simultaneously on multiple levels. We present a general approach to designing feedback control hierarchies for redundant biomechanical systems, that approximate the (non-hierarchical) optimal control law but have much lower computational demands. The approach is applied to the task of reaching, using a detailed model of the human arm. Our hierarchy has two levels of feedback control. The high level is designed as an optimal feedback controller operating on a simplified virtual plant. The low level is responsible for transforming the dynamics of the true plant into the desired virtual dynamics. The new method may be useful not only for modelling the neural control of movement, but also for designing functional electric stimulation systems that have to achieve task goals by activating muscles in real time.