As　robotic　technologies　advance,　robotic　arm　manipulation　tasks　in　complex　environments　become　increasingly　important.　This　paper　presents　a　new　collaborative　pushing　and　grasping　strategy　to　address　two　major　challenges　robotic　arms　encounter　in　complex　environments:　invisibility　and　tight　enclosure　of　target　objects.　We　propose　a　generator-evaluator　network　architecture　based　on　multiple　action　primitives,　incorporating　hierarchical　reinforcement　learning　and　Double　DQN　techniques　to　optimize　the　operational　efficiency　of　the　robotic　arm.　By　introducing　a　new　action　primitive　to　prioritize　pushing　operations　around　the　topmost　objects　in　space,　our　approach　effectively　reveals　and　locates　obscured　target　objects.　Additionally,　utilizing　the　Spatial　Correlation　Test　(SCT),　our　method　explores　all　possible　collision-free　operational　positions　in　the　action　space,　selecting　the　optimal　position　from　among　them.　Experimental　results　demonstrate　that　our　approach　achieves　high　efficiency　and　success　rates　in　scenarios　with　randomly　distributed　objects,　objects　tightly　enclosed　by　others,　and　invisible　objects.　This　strategy　enhances　the　speed　and　accuracy　of　tasks　and　improves　the　robot’s　adaptability　to　complex　environments.　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2025.