To　enable　robots　to　safely　grasp　target　objects　in　cluttered　environments,　a　precise　understanding　of　the　spatial　relationships　between　the　target　objects　and　their　surrounding　counterparts　is　essential.　While　convolutional　neural　networks　(CNNs)　demonstrate　potential　in　relational　reasoning,　their　primary　focus　on　pixel-level　feature　extraction　limits　their　ability　to　comprehend　the　global　context　and　critical　object　relationships,　subsequently　affecting　inference　accuracy.　To　address　these　limitations,　we　propose　a　relationship　reasoning　model　based　on　Graph　Attention　Networks　aimed　at　enhancing　the　accuracy　of　spatial　relationship　understanding　among　objects.　Initially,　we　employ　EfficientNet-BO　combined　with　a　Bidirectional　Feature　Pyramid　Network　(BiFPN)　for　RGB　feature　extraction　during　the　detection　process.　To　alleviate　the　computational　bürden,　we　filter　out　object　pairs　that　lack　clear　contextual　spatial　relationships.　We　then　utilize　a　sparsified　Graph　Attention　Network　that　incorporates　directional　attention　for　effective　relationship　reasoning.　The　proposed　model　is　trained　and　evaluated　on　the　Visual　Manipulation　Relationship　Dataset　(VMRD),　with　attention　visualized　using　Gradient-weighted　Class　Activation　Mapping　(Grad-CAM).　The　method　proposed　in　this　paper　was　evaluated　on　the　VMRD　dataset,　achieving　the　highest　precision　across　mAP,　OR,　and　IA　metrics.　This　reflects　an　improvement　in　both　object　detection　and　object　relationship　reasoning　tasks.　Specifically,　the　mAP　metric　reached　96.1%,　indicating　that　the　unique　BiFPN　structure　in　the　EfficientDet　network　better　integrates　features　of　different　scales　within　the　image,　effectively　enhancing　the　average　precision　of　image　detection.　The　significant　improvements　in　Object　Recall　(OR)　and　Image　Accuracy　(IA)　demonstrate　that　our　method　can　correctly　infer　a　greater　number　of　object　relationship　pairs　during　the　reasoning　phase.　Comparative　experiments　against　other　methodologies　on　the　same　dataset　reveal　that　our　model　significantly　improves　the　accuracy　of　relationship　reasoning,　demonstrating　its　applicability　and　extensibility　to　real　robotic　arm　grasping　scenarios.　The　model　achieves　an　image-based　accuracy　(IA)　of　71.　1%　in　relational　reasoning　tasks.　To　validate　the　effectiveness　of　the　proposed　model,　we　employed　a　technique　called　Gradient-weighted　Class　Activation　Mapping　(Grad-CAM),　which　is　used　to　interpret　the　decision-making　process　of　deep　convolutional　neural　networks.　Its　primary　aim　is　to　visualize　the　attention　distribution　of　the　neural　network　on　input　images　during　Classification　tasks,　aiding　in　the　understanding　of　the　model＇s　predictive　decision-making　process.　Grad-CAM　visualizations　further　substantiate　the　model＇s　capability　to　infer　spatial　relationships　among　multiple　objects　in　cluttered　scenes,　under-scoring　its　suitability　for　real-world　robotic　applications.　Additionally,　we　established　a　visual　grasping　experimental　platform　based　on　the　AUBO-i5　robotic　arm,　equipped　with　a　two-finger　electric　gripper　and　a　depth　camera.　To　validate　the　practical　application　and　generalization　ability　of　the　proposed　model,　we　constructed　a　specific　test　set　in　a　laboratory　environment.　The　collected　real　grasping　scene　examples　were　used　as　a　new　test　set　to　assess　the　model＇s　generalization　capability.　The　results　indicate　that　the　method　described　in　this　paper　still　exhibits　good　Performance　on　the　new　dataset.　While　our　RGB-based　Graph　Attention　Network　effectively　predicts　relationships　among　visible　objects,　it　is　validated　for　scenarios　involving　2　to　5　objects.　Future　research　will　focus　on　integrating　robotic　operational　actions　and　exploring　methods　to　infer　information　about　occluded　objects　based　on　the　positional　relationships　of　visible　objects.　We　will　also　investigate　strategies　to　enhance　model　Performance　in　scenarios　with　more　than　five　objects　and　conduct　physical　experiments　in　increasingly　complex　real-world　operational　environments　to　validate　effectiveness　and　identify　additional　areas　for　improvement.　©　2025　Science　Press.　All　rights　reserved.