Multi-robot Cooperation Research Based On Reinforcement Learning

The multi-robot system is the most challenging area of robotics research in recentyears. It includes a wide range of applications, such as mapping, multi-robotcooperative carrying and sports confrontation, etc. The multi-robot system canbrilliantly complete tasks that are difficult for a single robot through a cooperativemechanism. Today, people begin to pay more attention to this mechanism, to make themulti-robot system have the ability of adapting to unknown environment, one of themost promising ways to generate cooperative action is to let the robot interact with theenvironment and other robots. It will become one of the most potential methods inrobotics research.Reinforcement learning algorithm, a class of advanced machine learning methods,is developing rapidly in recent decades. It does not need priori knowledge, but to getknowledge through interaction with the environment, resulting in improved operationalstrategies and self-learning ability. In this article, lifting and carrying an object isconsidered as the task of multi-robot system, and researches mainly addressed oncooperative behavior acquisition based on reinforcement learning. The work done is asfollows:(1) In the article, joint-action is introduced to the traditional reinforcementlearning, and multi-robot reinforcement learning algorithm based on action predictionis presented to make multi-robot system own prediction mechanism thought. First, themodel of multi-robot prediction and reinforcement learning is constructed. Based onQ-learning, the state space and action space are rationally divided, and thereinforcement function is designed. Prediction function is used to reduced thedimension of reinforcement learning to speed up the convergence. Compared with nocooperation and traditional reinforcement learning algorithm, the simulationexperiments prove that action prediction can effectively facilitate cooperation.(2) In the article, Q-learning combines with BDI model, which makesmulti-robot system own logical reasoning capabilities. The behaviors of avoidingobstacles and moving to the target area have different weights, which automaticallyadjust with the environment. The evaluation function changing with location ofmulti-robot and distance from the last obstacle is presented. This innovative approachin the simulation experiment has achieved good results, and the multi-robot system caneasily complete the cooperative carrying.