In this paper techniques to identify and apply a restricted input subspace within the iterative learning control framework are developed. This is motivated by the increasing popularity of electrode arrays within rehabilitation and assistive technology communities, which allow functional electrical stimulation (FES) to be applied independently to each array element. This enables more selective muscle activation and improved control of human motion, but increases the input space dimension significantly so that model identification becomes impractical. The approach in this paper embeds past experience and/or structural knowledge in the subspace selection, and derives iterative learning controllers with favorable properties that are independent of the input basis employed. Experimental results using a 40 element surface electrode array confirm accurate tracking of three reference hand postures.