语义-几何协同引导的解耦式语言驱动六自由度抓取方法OA

To address the semantic and geometric feature entanglement caused by implicit coupling mechanisms,as well as the spatial misalignment between 2D visual perception and 3D geometric reasoning in language-driven 6-Degree-of-Freedom(6-DoF)grasping tasks within unstructured open scenarios,this paper proposes a decoupled language-driven grasping method guided by semantic-geometric synergy.The proposed method explicitly decouples the task into two syn-ergistic stages:cross-modal visual grounding and local geometric reasoning.In the visual grounding stage,a CLIP visual adapter is designed to introduce alignment priors,thereby achieving front-end cross-modal feature alignment.Simultane-ously,a bidirectional encoder is employed for iterative feature refinement,effectively enhancing the model's capability to parse complex instructions.In the geometric reasoning stage,a semantics-guided geometric cascade strategy is adopted.By leveraging high-confidence masks from the grounding stage for semantic-level cropping in the point cloud space,the pre-trained FGC-GraspNet is guided to generate grasp poses within the reconstructed local target point cloud space,effec-tively mitigating environmental noise interference.For the visual grounding task,the proposed model achieves intersec-tion over union(IoU)scores of 94.91%and 90.03%on the OCID-VLG and the seen-category test set of RoboReflt,respectively,outperforming HiFi-CS by 6.65 percentage points on OCID-VLG.Real-world robotic arm grasping experi-ments further validate the effectiveness of the decoupled framework in optimizing geometric inputs:in simple,complex,and challenging scenes,the system's grasping success rates for seen objects reach 97.5%,90.0%,and 90.0%,respectively,while the success rates for unseen objects reach 86.7%,73.3%,and 67.6%,respectively.Meanwhile,compared with existing end-to-end methods,the decoupled framework demonstrates higher robustness in multi-interference environments.Addi-tionally,the system achieves an inference speed of approximately 5.34 FPS with low computational overhead,validating its feasibility for deployment on edge devices.