Robot skill learning method and system based on reinforcement learning and unsupervised learning

Abstract

The invention discloses a robot skill learning method and system based on reinforcement learning and unsupervised learning, and the method comprises the steps: enabling a robot controller to control an execution mechanism to act when a specific task is executed, and enabling an image collection subsystem to collect a real-time state image of the execution mechanism in a task environment. An encoder of the unsupervised learning subsystem converts the real-time state image into a real-time state code, takes a state image when a set execution mechanism executes and completes a specified task as a target image, and converts the target image into a target code; the deep reinforcement learning subsystem outputs an action value of the execution mechanism at the next moment according to the real-time state code and the target code; and the robot controller controls the execution mechanism to make corresponding actions in real time according to the action value of the execution mechanism until the execution mechanism executes and completes the task. Compared with a common end-to-end network, the system provided by the invention is more universal, can be applied to different targets and tasks, and can also compress the training time.

Description

technical field [0001] The invention relates to the technical field of robot control, in particular to a robot skill learning method and system based on reinforcement learning and unsupervised learning. Background technique [0002] Reinforcement learning is a machine learning algorithm widely used at present. Its biggest feature is that it can learn the mapping from state to action from the environment. Based on the artificially designed reward function, a series of action decisions are based on maximizing Cumulative rewards are the goal. Among them, deep reinforcement learning is based on deep neural network, with state as network input and action as network output value, and the best decision-making network is obtained through continuous training. Deep reinforcement learning can be applied to various scenarios that require decision optimization, such as game AI and robot control. [0003] With the development of robot technology, the application scenarios of robots are ...

AI Technical Summary

Problems solved by technology

However, using an end-to-end deep neural network is generally difficult to converge and requires more training time, and for different tasks, different reward functions need to be designed. The specially customized network structure and parameters make it only complete specific tasks, which is difficult Expand to other different tasks

Images