A robot grasping detection method and system

Abstract

The present application is a robot grasping detection method and system, the method mainly includes: (1) identifying the object category; (2) searching for the unknown object on the first public data set to obtain the marked data, and updating the object category judgment model; (3) ) Update the grasping detection model, and re-detect the grasping frame of the object; (4) Control the robot to actually grasp the object. When the object is identified as unknown, retrieve the marked data in the public data set and update the grasping detection model, without additional simulation environment and manual intervention, effectively improving the grasping ability of the grasping detection model for unknown objects; using known objects The model is updated with the image samples of unknown objects, and the model training method of incremental learning is adopted, which can dynamically update the knowledge base of object grasping, maintain the memory ability of the grasping detection model for the original marked data and improve the ability of the new marked data learning ability.

Description

Technical field [0001] The present invention belongs to the field of robot vision systems capture control, particularly to a method and system for detecting robotic gripping. Background technique [0002] Traditional robots to crawl frame detection method is based on a closed world, the training data contains only known object tag data, grab the box after good detection model training data for training, knowledge base will set the type, follow-up can not dynamically update the knowledge base crawl detection model thus obtained on the identification of a known object capable of producing good results; however, lack the marking data of the unknown object, gripping detection model training and when there is no corresponding data, the unknown identification of objects, the robot can not generate good gripping block. [0003] The practical application of previously collected tag data of all objects is not realistic, existing technical solutions for the problem of insufficient training...

AI Technical Summary

Problems solved by technology

[0002] The traditional robot grabbing frame detection method is based on a closed world. The training data only contains the labeled data of known objects. After the grabbing frame detection model is trained for the training data, the knowledge base is finalized, and the knowledge base cannot be dynamically updated in the future. , the grasp detection model obtained in this way can produce good results in the recognition of known objects; but due to the lack of labeled data for unknown objects, the grasp detection model does not have corresponding data during training, so for unknown Object recognition, the robot cannot generate a good grab frame

[0003] However, it is unrealistic to collect the labeled data of all objects in advance in practical applications. To solve the problem of insufficient training samples, the existing technical solutions mainly include the method of generating virtual data combined with the simulation environment, semi-supervised method, self-supervised method, etc. These methods It still relies on human intervention and environmental interaction, and fails to make effective use of existing public grasping datasets, which often contain labeled data of unknown objects. The time cost and computing cost of existing technical solutions increase, and the model training less efficient

Images