Traditional　muscle　strength　training　instruments　often　rely　on　torque-based　feedback　to　guide　exercises,　which　can　introduce　delays　in　system　response　and　result　in　discomfort　due　to　hysteresis　effects.　Surface　electromyography　(sEMG)　signals　were　used　as　control　inputs　to　overcome　the　lag　in　torque-based　muscle　strength　training　instruments.　The　sEMG　is　generated　20-80　ms　before　movement,　which　is　called　＂muscle　electromechanical　delay.＂　If　the　torque　can　be　effectively　predicted　during　this　period,　the　lag　effect　can　be　significantly　reduced,　thus　improving　the　effectiveness　and　comfort　of　training.　We,　therefore,　proposed　a　multistep　ahead　(MSA)　model　based　on　the　nonlinear　autoregressive　network　with　exogenous　inputs　(NARX)　dynamic　recurrent　neural　network.　It　predicted　torques　using　sEMG,　and　allowed　natural　control　of　the　instrument.　The　results　showed　that　the　normalized　root-mean-square　error　(NRMSE)　was　lower　than　0.1167,　and　the　Pearson　correlation　coefficients　(rho)　exceeded　0.9444,　even　when　the　ahead　steps　achieved　35.　The　intrasubject　and　the　intersubject　validation　demonstrated　significantly　lower　NRMSE　(　p　＜0.05)　and　higher　rho　(　p　＜　0.05)　of　the　MSA　model,　compared　with　some　state-of-theart　recursive　models　and　typical　models　without　autoregression　items.　It　proves　that　the　MSA　can　accurately　predict　the　motion.　Meanwhile,　the　introduction　of　the　sEMG　signal　as　a　control　source　significantly　reduced　the　root-mean-square　jerk　(RMSJ)　of　the　torque,　demonstrating　smoother　motion.　The　experimental　results　revealed　that　the　one-step-ahead　model　achieved　an　average　response　time　of　3.73　ms,　which　is　markedly　lower　than　the　muscle　electromechanical　delay.　The　response　time　increased　by　an　average　of　approximately　0.068　ms　per　additional　ahead　step.　In　conclusion,　the　proposed　EMG-driven　muscle　strength　training　instrument　enables　natural　muscle　strength　training.