Learning Based Autonomous Robot Decision Making Technology

Robots in complex environment are usually characterized by continuous interaction with the environment,autonomy,and adaptability,which requires the robot to make decisions autonomously in combination with environmental conditions and tasks.Robot decision-making needs to be true and effective to promote the robot to achieve the mission objectives,to adapt to changes in the environment and should try to meet real-time requirements.Therefore,how to provide effective methods to deal with robot's decision-making in complex environment is an important research topic.Reinforcement learning methods can make decision-making learning from the process of interacting with the environment,while deep learning can effectively extract the feature information of high-dimensional data.The use of learning technology to solve the problem of robot decision-making has become an important method.Prior knowledge is the information about tasks and environment obtained by robots before making decision learning.Therefore,using priori knowledge can improve the speed and effect of robot decision learning.In order to overcome the complex environment for robot decision-making problems,this paper on how to use machine learning techniques and prior knowledge to assist the robot decision-making modeling,optimization and evaluation.In this paper,we demonstrate the effectiveness of robot decision learning through experiments on repeatability in different environments.The main achievements of this paper and innovation are summarized as follows.Aiming at the problem of robot decision-making under the state space of high-dimensional environment,a decision learning algorithm based on reinforcement learning is designed and implemented.This paper abstract the robot's decision-making process into a series of discrete events that the robot interacts continuously with the environment.Based on the reinforcement learning model design,a learning-based robot decision-making learning algorithm is realized,which can effectively accumulate from high dimensional environment state space Experience in decision-making learning.Aiming at the problem of blindness and repeatability in the exploration of environmental state space,an environment state space exploration algorithm based on sample data is proposed.In order to overcome the repeatability and blindness in the exploration of environmental state space,this paper draws lessons from the way human beings can expand learning based on the example data and further introduces the example data into the robot decision-making learning model,and proposes a ripple exploration Strategy and state-space exploration algorithm based on sample data: EX-D to explore the highdimensional state space of the environment.Experiments based on MountainCar and Breakout task environment are carried out to verify the effectiveness of the above algorithm.In this paper,a standardized test platform is used to analyze and experiment on different environments in order to demonstrate the superiority of learning-based robots compared to traditional decision-making methods and the optimization effect of prior knowledge on robot decision-making learning process.Experimental data shows that under certain circumstances,learning-based robot decision making algorithm is more effective than traditional robotic decision making techniques,while using EX-D exploration algorithm can improve decision learning by more than 25%,and also improve the convergence speed of the algorithm.