In　order　to　increase　the　computational　efficiency　of　neural　networks,　a　new　network　model　named　state　delay　input　dynamic　recurrent　neural　network　is　presented.　This　new　neural　network　is　applied　to　the　model　identification　of　PowerCube　modular　robot　system　with　all　kinds　of　disturbing　factors.　The　data　of　joint　positions　retrieved　from　the　robot　and　the　position　of　the　end-effector　measured　by　the　OPTOTRAK　3020　are　used　as　learning　sets　for　neural　network.　The　learning　superiority　of　the　new　neural　network　is　illustrated　and　the　validity　of　neural　network　models　for　robot　joints　are　proved　by　inputting　validating　sets　and　analyzing　experimental　results　and　errors.