A robot grasping detection method and system for cluttered scenes

By introducing an encoder-processor-decoder architecture and KA-Cross attention blocks into robot grasping and detection, the problem of poor generalization in cluttered scenes is solved, achieving higher accuracy and automated grasping processes.

CN121083672BActive Publication Date: 2026-03-03HUNAN XIANGJIANG TIMES ROBOT RES INST CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511648631.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-11-12
Publication Date
2026-03-03
Estimated Expiration
2045-11-12

AI Technical Summary

Technical Problem

Existing technologies have poor generalization ability in robot grasping and detection in cluttered scenarios and are difficult to adapt to complex input information, resulting in low grasping and detection accuracy.

Method used

A robot grasping and detection network model based on an encoder-processor-decoder architecture is adopted, which includes KA-Cross attention blocks. The cross attention blocks introduce induction points to enhance the recognition and screening capabilities of key image features. The model is trained and deployed using deep learning technology.

Benefits of technology

It improves the accuracy and success rate of robot grasping and detection, can adapt to complex environments, reduce human intervention, realize fully automated processes, and improve work efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121083672B_ABST
    Figure CN121083672B_ABST
Patent Text Reader

Abstract

This invention relates to the field of robot object grasping, specifically a method and system for robot grasping and detection in cluttered scenes. The method includes: 1. Acquiring RGB and depth images of objects in a cluttered scene, preprocessing them, and constructing a training dataset; 2. Building a robot grasping detection network model; 3. Selecting RGB and depth images from the training dataset and inputting them into the robot grasping detection network model to obtain predicted grasping postures; 4. Constructing a loss function based on the predicted and actual grasping postures, repeating steps 3 and 4 to minimize the loss function until it converges; 5. Deploying the trained robot grasping detection network model to a robot platform, using the model to generate the object's grasping posture, and completing the grasping of the target object. This invention can accurately locate target objects in cluttered scenes and predict the optimal grasping point, thereby effectively improving the robot's grasping success rate.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Human-computer interaction behavior evaluation method based on KAN-Transform and fuzzy reasoning

    CN119046818A

  • KAN-based adversarial unsupervised time sequence anomaly detection method and system

    CN120257171A