Prediction　of　actively　exerted　torque　from　the　ankle　joint　complex　(AJC)　is　critical　for　human-robot　interaction　because　this　is　the　real　torque　that　the　patient　exerts.　In　this　work,　we　develop　a　prediction　framework　to　predict　the　actively　exerted　torque　from　the　surface　electromyography　(sEMG)　signal.　The　modified　non-negative　matrix　factorization　(mNMF)　algorithm　is　used　to　extract　muscle　activation　information　(MAI)　and　muscle　synergy　information　(MSI);　the　pseudo-siamese　network　(PSN)　is　used　to　fit　the　MAI　and　MSI　to　actively　exerted　torque.　The　prediction　accuracy　is　compared　under　different　ankle　attitudes,　synergistic　feature　extraction　algorithms,　and　prediction　algorithms,　and　the　results　show　that　the　prediction　accuracy　of　the　proposed　framework　is　94.15%,　and　the　variance　of　prediction　accuracy　is　just　3.56%　under　different　ankle　attitudes.　The　proposed　framework　can　provide　the　real　actively　exerted　torque　for　the　active　rehabilitation　training　of　ankle　joint,　so　as　to　improve　the　enthusiasm　and　initiative　of　patients　during　rehabilitation　training.　IEEE