Joint　torque　prediction　plays　an　important　role　in　quantitative　limb　rehabilitation　training　and　the　exoskeleton　robot.　The　Surface　electromyography　signal　(sEMG)　with　the　advantages　of　non-invasive　and　easy　collection　can　be　applied　to　the　prediction　of　human　muscle　force.　By　utilizing　the　sEMG,　the　recurrent　cerebellar　model　neural　network　(RCMNN),　which　has　better　generalization　and　computational　power　than　the　traditional　neural　network　has　been　used　to　predict　the　joint　torque.　In　this　work,　a　smooth　function　with　adaptive　coefficient　is　employed　to　polish　the　results　of　RCMNN,　the　proposed　method　shows　great　performance　on　torque　prediction　with　the　correlation　coefficient　between　the　torque　and　the　estimation　result　up　to　98.43%,　such　advanced　model　paves　the　way　to　the　application　on　the　quantitative　rehabilitation　training.