Purpose　-　The　safety　performance　of　cooperative　robots　is　particularly　important.　This　paper　aims　to　study　collision　detection　and　response　of　cooperative　robots,　which　meet　the　lightweight　requirements　of　cooperative　robots　and　help　to　ensure　the　safety　of　humans　and　robots.　Design/methodology/approach　-　This　paper　proposes　a　collision　detection,　recognition　and　response　method　based　on　dynamic　models.　First,　this　paper　identifies　the　dynamic　model　of　the　robot.　Second,　an　external　torque　observer　is　established　based　on　the　model,　and　a　dynamic　threshold　collision　detection　method　is　designed　to　reduce　the　interference　of　model　uncertainty　on　collision　detection.　Finally,　a　collision　position　and　direction　estimation　method　is　designed,　and　a　robot　collision　response　strategy　is　proposed　to　reduce　the　harm　caused　by　collisions　to　humans.　Findings　-　Comparative　experiments　are　conducted　on　static　threshold　and　dynamic　threshold　collision　detection,　and　the　results　showed　that　the　static　threshold　only　detected　one　collision　while　the　dynamic　threshold　could　detect　all　collisions.　Conducting　collision　position　and　direction　estimation　and　collision　response　experiments,　and　the　results　show　that　this　method　can　determine　the　location　and　direction　of　collision　occurrence,　and　enable　the　robot　to　achieve　collision　separation.　Originality/value　-　This　paper　designs　a　dynamic　threshold　collision　detection　method　that　does　not　require　external　sensors.　Compared　with　static　threshold　collision　detection　methods,　this　method　can　significantly　improve　the　sensitivity　of　collision　detection.　This　paper　also　proposes　a　collision　position　direction　estimation　method　and　collision　separation　response　strategy,　which　can　enable　robots　to　achieve　post　collision　separation　and　improve　the　safety　of　cooperative　robots.