Research on Target Detection and Localization System for Robotic Grasping Based on YOLOv11

Robotic grasping performance in unstructured environments fundamentally depends on visual perception. This paper presents a detection-localization for NAO robot grasping, integrating an improved YOLOv11 network with Gaussian Process Regression (GPR)-based error compensation. The enhanced detector incorporates depthwise separable convolutions for lightweight architecture (35.7% parameter reduction, 83 FPS), a convolutional block attention mechanism for feature refinement, and adaptive multi-scale fusion for small-object sensitivity. The GPR module models depth residuals using detection confidence and bounding box geometry, suppressing systematic monocular ranging errors. Experimental results show the improved detector achieves 95.3% [email protected], a 4.1% gain over baseline YOLOv11. Following compensation, ranging RMSE decreases from 3.87 cm to 0.98 cm globally, with 69.9% error reduction in the 90–120 cm range. The integrated system attains grasping success rates of 98.0%, 92.0%, and 85.0% under no occlusion, partial occlusion, and stacking scenarios, with 219 ms. This work provides an effective solution for high-precision real-time perception on resource-constrained platforms.