Robot operating system and method based on video incremental learning

Abstract

The invention relates to a universal technology in the field of robots, and particularly discloses a robot operating system and method based on video incremental learning. The robot operating method comprises the steps that S1, a target positioning model is constructed, and pre-training processing is conducted on an MS-COCO data set; S2, an increment classification model is constructed, and pre-training processing is conducted on an ImageNet data set; S3, incremental training is conducted on the existing models based on a new target; and S4, the position of the selected target is positioned, and the grabbing movement track of a mechanical arm is constructed. According to the system and method, based on a detection-classification two-stage normal form, different from a general classification problem, a classification task based on video information has additional time information, different target postures are beneficial to constructing a better classification module, and the problem of low-resolution recognition under unfavorable conditions such as darkness and the like can be solved.

Description

technical field [0001] The invention relates to general technology in the field of robots, in particular to a robot operating system and method based on video incremental learning. Background technique [0002] In the field of machine learning, incremental learning is dedicated to solving a common flaw in model training: catastrophic forgetting, that is, general machine learning models (especially back-propagation-based deep learning methods) fail to perform well on new tasks. Performance on old tasks often drops significantly when training on . The actual application scenarios of robot grasping technology are complex and diverse. Therefore, in the dynamic open scene, the key challenge for the robot to accurately identify and grasp the correct object is to deal with the uncertainty and constant change in the open environment. [0003] In order to make the robot adapt to the dynamic open environment, the invention enables the robot to have the ability of progressive learning...

AI Technical Summary

Problems solved by technology

[0002] In the field of machine learning, incremental learning is dedicated to solving a common flaw in model training: catastrophic forgetting, that is, general machine learning models (especially back-propagation-based deep learning methods) fail to perform well on new tasks. Performance on old tasks often drops significantly when training on

However, the object features captured by these methods are either one-sided or too comprehensive, and there is a large deviation from the actual object features in the real scene.

In addition, the processing method for new objects in the existing technology is too complicated. Whether it is to provide static pictures from different perspectives or to build a 3D model of the object, the time cost or calculation cost is too high, which is not conducive to the application and promotion of this technology.

Images