In this talk, we will consider nonparametric regression models with multiple functional covariates. The focus of the work is to identify relevant variables and useful metrics for the functional covariates, and to efficiently estimate the regression function. The proposed method is based on an extension of the Nadaraya-Watson estimator, where a kernel function is applied to a linear combination of distance measures, each computed on individual covariates, in combination with an adaptive thresholding step on the kernel weights. This data-driven least squares cross-validation method can asymptotically remove irrelevant noise variables and select relevant metrics for the functional covariates, as will be shown both by theory and numerical examples.