Robotic hand prostheses translate forearm electromyography (EMG) signals into finger and wrist movement through trained pattern recognition (PR) algorithms. Two major limitations of this approach are the need for extensive training and recalibration across different arm postures. We develop a novel EMG-based functional linear model that accounts for the underlying biomechanics of hand movement, leading to natural, continuous movement of a robotic hand prosthetic that is not position dependent. The model is made parsimonious and interpretable through a two-step variable selection procedure motivated by the relaxed LASSO technique. A final model is then fit on the selected subset of EMG signals to reduce shrinkage bias of the regression functions. Our variable selection method is shown to identify clinically important EMG signals with negligible false positive rates for an able-bodied subject and, unlike PR, holds across different postures. The proposed methodology is also applicable in a general functional linear model setting where the functional coefficients vary over multiple covariates. An extensive simulation study shows excellent variable selection and predictive performance.