基于深度学习的工业机器人抓取识别方法及系统
By constructing a high-fidelity point cloud in collaboration with structured light polarization imaging and a force sensor, and combining PointNet++ and U-Net networks, the problems of visual data degradation and missing physical properties in the scenario of closely stacked highly reflective metal gears are solved, and a grasping decision with high reliability and stability is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JIANGSU TINGBEI TECH CO LTD
- Filing Date
- 2025-11-07
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies suffer from visual data degradation, incomplete information, and decision ambiguity in scenarios involving closely stacked highly reflective metal gears, making it difficult to guarantee the reliability and stability of grasping decisions.
The initial depth map is acquired using structured light polarization imaging, and the three-dimensional contact points are received by a force sensor to generate a high-fidelity gear point cloud. The centroid position probability and friction coefficient distribution are calculated using the PointNet++ architecture to construct the action value function. The pixel-level grasping and pushing action value map is generated using U-Net to form a joint strategy space for adaptive decision-making.
It achieves highly reliable gripping in scenarios with closely stacked highly reflective metal gears, ensuring the geometric feasibility and physical stability of gripping, improving gripping accuracy and stability, and preventing slippage or displacement.
Smart Images
Figure CN121415090B_ABST