Robot　imitation　learning　has　recently　been　studied　extensively.　There　have　been　many　methods　for　robot　imitation　learning.　The　main　problem,　however,　is　that　the　learned　strategy　is　susceptible　to　external　interference　and　thus　deviates　front　the　expected　trajectory.　In　this　paper　we　propose　a　method　for　robot　learning　human　motion　that　combines　the　strengths　of　LSTM　and　dynamical　system.　Such　an　approach　can　learn　policy　effectively　from　dynamical　system　of　motions　while　robot　interacting　with　human　demonstrators　or　other　teachers.　And　this　method　effectively　improves　the　ability　of　robot　trajectory　adjustment.　Simulation　experiments　show　that　the　proposed　method　which　learn　policy　can　effectively　avoid　the　disturbance　in　the　robot　working　environment.