In　this　paper,　we　propose　a　3D　positioning　framework　based　on　convolutional　neural　network　to　realize　the　pose　estimation　of　the　target　object　during　the　mechanical　arm　grabbing　process.　First,　the　image　data　for　training　this　network　is　completely　synthesized　by　domain　randomization　technology　to　reduce　the　data　acquisition　cost　and　make　up　the　gap　between　the　simulated　image　and　the　real　image.　Then,　the　3D　positioning　framework　is　designed　based　on　the　ResNet　architecture,　which　consists　of　two　parts.　One　is　to　obtain　the　classification　of　target　object,　and　the　other　is　to　obtain　the　3D　coordinates　of　the　object　and　the　horizontal　rotation　angle.　Finally,　the　target　image　classification　and　pose　estimation　are　tested　on　the　real　images　with　interference　and　the　synthetic　images　with　interference,　respectively.　The　results　show　that　the　proposed　method　has　high　prediction　accuracy,　and　the　effectiveness　of　the　framework　is　proved　by　migrating　it　to　the　real　physical　robot　arm　experimental　platform.　©　2019　IEEE.