Function-on-scalar　regression　is　commonly　used　to　model　the　dynamic　behaviour　of　a　set　of　scalar　predictors　of　interest　on　the　functional　response.　In　this　article,　we　develop　a　robust　variable　selection　procedure　for　function-on-scalar　regression　with　a　large　number　of　scalar　predictors　based　on　exponential　squared　loss　combined　with　the　group　smoothly　clipped　absolute　deviation　regularization　method.　The　proposed　procedure　simultaneously　selects　relevant　predictors　and　provides　estimates　for　the　functional　coefficients,　and　achieves　robustness　and　efficiency　using　tuning　parameters　selected　by　a　data-driven　procedure.　Under　reasonable　conditions,　we　establish　the　asymptotic　properties　of　the　proposed　estimators,　including　estimation　consistency　and　the　oracle　property.　The　finite-sample　performance　of　the　proposed　method　is　investigated　with　simulation　studies.　The　proposed　method　is　also　demonstrated　with　a　real　diffusion　tensor　imaging　data　example.