Continuous　motion　estimation　based　on　electromyographic　signals　(sEMG)　typically　confronts　challenges　posed　by　physiological　variances　among　subjects.　These　variances　frequently　render　it　difficult　for　the　model　to　achieve　satisfactory　generalization　across　different　subjects.　For　this　purpose,　this　paper　proposes　a　method　integrating　adversarial　transfer　learning　(ATL)　and　maximum　mean　discrepancy　(MMD).　This　approach　effectively　aligns　the　feature　distributions　among　different　subjects.　Experimental　outcomes　indicate　that,　in　comparison　to　advanced　transfer　learning　methods,　the　proposed　method　realizes　remarkable　enhancements　in　both　prediction　accuracy　and　model　stability,　laying　a　solid　foundation　for　the　application　of　inter-subject　continuous　motion　estimation.　©　2024　Copyright　held　by　the　owner/author(s).　Publication　rights　licensed　to　ACM.