In　order　to　improve　the　safety　performance　of　the　robot,　this　paper　proposes　an　external　force　detection　method　for　humanoid　robot　arm　without　using　joint　torque　sensors,　which　can　detect　the　external　force　of　the　joint　space　in　real　time　during　the　operation　of　the　robot.　First,　Analyses　on　the　structure　of　the　humanoid　robot　arm　is　performed,　and　the　model　of　robot　external　force　detection　is　established　based　on　robot　dynamics　and　motor　dynamics.　Then,　the　error　of　detection　model　is　analyzed,　and　the　robot　dynamic　model　error　is　compensated　by　using　the　BP　network,　in　order　to　obtain　more　accurate　external　force　detection　value　of　the　robot.　Experiments　show　that　the　method　can　effectively　improve　the　detection　accuracy,　and　the　obtained　external　force　detection　data　can　be　applied　not　only　to　the　collision　detection　in　the　static　state　of　the　robot,　but　also　to　the　online　collision　detection　when　the　robot　is　running,　which　can　ensure　the　safe　operation　of　the　robot.