Disordered grabbing multi-objective optimization method and system based on deep reinforcement learning

Abstract

The invention belongs to the field of disordered grabbing of mechanical arms, and particularly relates to a disordered grabbing multi-objective optimization method and system based on deep reinforcement learning, the disordered grabbing multi-objective optimization method based on deep reinforcement learning processes the same scene at the same moment through two parallel and independent Q networks, and the mechanical arm grabs the respective grabbing points of the two networks and returns parameters such as an execution path and grabbing power consumption. Between the Q networks, advantages and disadvantages of the two are distinguished about capture effects of an execution path, capture power consumption and the like, and corresponding reward values are generated. The Q network receives internal and external reward function feedback, the problem that a reward value function of a single Q network can only be discrete data is solved, and continuous data such as an execution path and grabbing power consumption are added into the reward value function, so that selection of grabbing points is further optimized.

Description

technical field [0001] The invention relates to the field of disordered grasping of a robotic arm, in particular to a multi-objective optimization method and system for disordered grasping based on deep reinforcement learning. Background technique [0002] With the development of robot technology, the application scenarios of the existing robot disordered grasping technology continue to expand, and the reinforcement learning method based on the grasping success rate as the network training target cannot effectively meet the multi-index differentiation of robots' disordered grasping in different application scenarios need. The efficient multi-objective optimization of the robot's disorderly grasping behavior has important practical significance for improving the robot's customized work ability and expanding the robot's application scenarios. [0003] The deep reinforcement learning algorithm has obvious intelligence and robustness, based on the feedback of the environment, t...

AI Technical Summary

Problems solved by technology

[0004] The reward value functions accepted by the existing Q network are all discrete, that is, the action execution results are divided into different situations according to the threshold and different rewards are given. Such reward feedback is suitable for pre-defined situations. In the process of target grasping, many factors that affect the grasping effect are continuously changing quantities, such as the grasping path, the power consumption of the robotic arm, etc. It is difficult to predict the effect of these variables in advance, so it is impossible to pre-determine the reward value for different situations how should change

Images