ROBOT MOTOR-SKILL LEARNING METHOD AND SYSTEM

WIPO Patent Application WO/2022/105635

A1

Disclosed are a robot motor-skill learning method and system, said method comprising: obtaining a human drag demonstration data sample set; performing dimensionality reduction on said data sample set on the basis of principal component analysis; establishing variable constraint conditions in a latent space and, in combination with said variable constraint conditions, filtering the data sample set after dimensionality reduction to generate a latent space data set; using a Gaussian mixture model with Gaussian mixture regression to perform modeling and learning of said latent space data set, to output a robot motion-control training model; performing prediction on said robot motion-control training model on the basis of a recurrent neural network to solve for the optimal solution of the model, and converting the model optimal solution into an actual control variable of the robot. In the embodiments of the present invention, autonomous learning of robot motor skills can be achieved by means of using a small amount of human demonstration data while taking into account the inherent constraints of the robot body, effectively improving the generalization ability and programming efficiency of an algorithm.

CHENG TAOBO (CN)
SU ZERONG (CN)
XU ZHIHAO (CN)
WU HONGMIN (CN)
LI XIAOXIAO (CN)
ZHOU XUEFENG (CN)

PCT/CN2021/129342

May 27, 2022

November 08, 2021

Click for automatic bibliography generation  

INSTITUTE OF INTELLIGENT MFG GUANGDONG ACADEMY OF SCIENCES (CN)

B25J9/16

CN112605973A	2021-04-06
CN109702744A	2019-05-03
CN108656119A	2018-10-16
CN110682286A	2020-01-14
CN110977965A	2020-04-10
KR20130067345A	2013-06-24

GUANGDONG GUANGYING INTELLECTUAL PROPERTY OFFICE (CN)

View/Download PDF      PDF Help

Previous Patent: COMMUNICATION METHOD AND COMMUNICATION DEVICE IN WIRELESS LOCAL AREA NETWORK

Next Patent: SPIRO COMPOUND, PHARMACEUTICAL COMPOSITION CONTAINING SAME, AND USE THEREOF