In　order　to　simplify　the　complex　motion　planning　and　improve　the　intelligence　of　robot　arm,　a　robot　arm　task　imitation　system　based　on　RNN　(Recurrent　Neural　Network)　is　proposed.　Firstly,　the　original　task　is　demonstrated　to　robot　arm,　and　the　original　data　is　collected　which　includes　original　task　trajectory　data　and　robot　arm　joint　angle　data.　Secondly,　RNN　is　constructed　and　used　to　obtain　imitation　policy　by　training　original　data.　Thirdly,　when　task　changes,　new　data　is　collected　which　only　include　new　task　trajectory　data,　and　robot　arm　joint　angle　data　is　obtained　by　imitation　policy　generalization　of　new　data.　The　experimental　results　show　that　the　imitation　system　not　only　can　simplify　complex　motion　planning　and　reproduce　demonstration　of　original　task,　but　also　can　realize　new　task　imitation　by　policy　generalization　when　task　changes.　©　2017　IEEE.