Reinforcement learning awarding method suitable for movable mechanical arm

Abstract

The invention discloses a reinforcement learning awarding method suitable for a movable mechanical arm. The method comprises the following steps that S1, a task scene, an initialization algorithm, various parameters of the movable mechanical arm and a deep network model corresponding to the algorithm are designed; S2, environment information sent by information collection equipment on the movablemechanical arm is reconstructed and rasterized, and a starting position and a target position of the movable mechanical arm are clear and definite; S3, interaction with the environment, training dataare collected and stored in an experience pond; S4, one batch of data is sampled from the experience pond, and through treatment of a reward function, rethinking rewards which are additionally designed are obtained for subsequent training; S5, original rewards and extra rewards are combined, and a deep reinforcement learning algorithm trains the movable mechanical arm to finish a target task in aplanned space; and S6, related training data and final model parameters obtained through training are recorded, and a corresponding optimal strategy is obtained.

Description

technical field [0001] The invention relates to a reinforcement learning reward method suitable for mobile manipulators. Reinforcement learning is used in traditional manipulator control tasks, and the structure of the reward signal is redesigned to solve the defects of the reward setting in the existing method, and achieve better results. Good control performance. Background technique [0002] In the traditional control field, the control and motion planning of the mobile manipulator usually involves kinematic modeling of the manipulator and solving the end pose and the angle value of each joint. As the complexity of the actual application environment continues to increase, the traditional model-based and rule-based control methods will become more and more complex in modeling the environment and solving the kinematics of the manipulator, and require a lot of human experience to design the corresponding rules. [0003] Both design rules and modeling require a lot of energy...

AI Technical Summary

Problems solved by technology

As the complexity of the actual application environment continues to increase, the traditional model-based and rule-based control methods will become more and more complex in modeling the environment and solving the kinematics of the manip

Images