Transfer learning method based on scene template generation

Abstract

The invention relates to a migration learning method based on scene template generation. The method comprises the steps of designing a scene template, constructing the scene template, training a template generation network and migrating a strategy network based on the scene template. According to the method, the unified template of different scenes is established through the concept of the scene template, the common characteristics of the different scenes are constructed, and the relationship between the different scenes is established through the common characteristics, so that the migrationof the strategy network obtained through reinforcement learning between the environments with relatively large differences is realized.

Description

technical field [0001] The invention relates to a migration learning method based on scene template generation, which belongs to the field of intelligent robots, and is used to realize the seamless migration of the strategy network trained by the robot in the mathematical simulation environment to the real physical environment, and to migrate between different physical environments to realize Skills learned in one setting can be applied in another. Background technique [0002] Research on "self-learning" space operation robots based on deep reinforcement learning methods refers to endowing robots with self-learning capabilities, and training robots to master various operating skills through "imitation" and autonomous training, so as to solve the unstructured environment that plagues current space operations , teleoperation and long delay are important research directions in the field of robotics. [0003] Robot self-training requires robots to learn operational skills thro...

AI Technical Summary

Problems solved by technology

[0007] The above two methods have great limitations in the migration process

Images