SeeGrasp: Semantic-Empowered Efficient Grasping via Robotic Active Spatial Perception

Abstract

Achieving robust robotic grasping in cluttered environments remains a significant challenge. While end-to-end learning methods are constrained by heavy data requirements and insufficient precision, traditional perceive-then-plan paradigms often suffer from passive perception and mismatches between geometric poses and optimal functional grasp points. To bridge these gaps, we propose SeeGrasp, a novel framework leveraging advanced foundation models. SeeGrasp operates in a trainingfree paradigm, requiring no data collection or retraining while maintaining high robustness. It incorporates a semantic-enhanced active perception strategy using multi-view scanning to overcome the limitations of passive observation. Furthermore, a semantic and 3D-guided functional grasping module is introduced to align estimated geometric poses with practical functional requirements. Real-world experiments yield an 85.8% average success rate, showing improvements over current state-of-the-art methods and highlighting its substantial potential for high-precision industrial and service manipulation.