Robot grabbing data set establishing device and method based on teaching and deep learning

Abstract

The invention discloses a robot grabbing data set establishing device and method based on teaching and deep learning, and relates to the field of robots. The method comprises the steps: 1, adopting a demonstrator for demonstrating grabbing poses on a target object, extracting grabbing part point clouds from a scene according to grabbing space, and carrying out manual scoring and label data labeling while data are collected; 2, rapidly training a six-degree-of-freedom grabbing pose evaluation network by using the collected label data; 3, quickly collecting point cloud data of the actually captured scene by adopting a continuous frame shooting mode; and 4, sampling six-degree-of-freedom grabbing poses on the actually grabbed scene point cloud, and evaluating and scoring each sampling pose by using the trained six-degree-of-freedom grabbing pose evaluation network so as to complete automatic generation of a label of the scene point cloud. According to the method, the technical difficulty that the hardware cost and the time cost for generating and establishing the six-degree-of-freedom grabbed data set in the past are relatively high is solved.

Description

technical field [0001] The present invention relates to the field of robots, in particular to a device and method for establishing a robot grasping data set based on teaching and deep learning. Background technique [0002] The disorderly grasping of arbitrary objects by robots is a classic research problem in the field of robotics. According to the input of visual sensors, robots predict the grasping pose of six degrees of freedom in the action space. Robots generally use two-finger grippers to grasp objects. action. Most of the prediction methods of the six-degree-of-freedom grasping pose are data-driven. Through the end-to-end neural network model, the sensor information in the form of point cloud is input, and the prediction of the grasping pose is output. [0003] The pose estimation algorithm for grasping any object based on machine learning is often data-driven and requires a large amount of labeled data to train the network model. Building a data set that can be us...

AI Technical Summary

Problems solved by technology

[0010] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is that the time cost and time cost of real data collection are relatively high, the migration of simulation data to the real environment is poor, and the network performance is degraded

Images