Font Size: a A A

Intelligent Assembly System Based On Multi-Agent Reinforcement Learning

Posted on:2021-01-09Degree:MasterType:Thesis
Country:ChinaCandidate:Y J LiFull Text:PDF
GTID:2518306308975429Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
Under the background of Industry 4.0 and China Manufacturing 2025,intelligent manufacturing and intelligent factory have become key targets that affect productivity levels and the national economy.The intelligent assembly system usually refers to a multi-agent system based on collaborative robotic arms for assembly tasks.Compared with the operation of a single-arm robot,cooperative robotic arms have more complicated motion interference problems.Therefore,traditional method is hard to meet the demands of various modern industries due to the huge amount of calculation.Multi-agent reinforcement learning provides a new solution for solving the problem of robotic arm assembly.Through autonomous exploration and policy learning of agents,the cooperation task of multi-agents is completed,which has high robustness and practical value.This paper proposes a cooperative robotic arms assembly system based on multi-agent reinforcement learning,aimed to make the robotic arms cooperate with each other to complete different types of assembly tasks through an end-to-end learning process.For the collaborative robotic arm assembly tasks,this paper will propose an improved MADDPG algorithm which introduce LSTM and curiosity mechanisms to solve the problem of POMDP and sparse rewards.For experiment process,this paper will use the Gym development tool and MuJoCo physics engine to establish a simulation environment for multi-agent systems and implement a segmented decomposition of the reward function to improve learning efficiency.Then,the improved MADDPG algorithm will be tested in simulation environment to verify the effectiveness of the algorithm.Experiments proves that the improved MADDPG has achieved ideal results in two of the three scenarios of collaborative assembly problem.And compared with original algorithm,the reward of improved MADDPG has respectively imporved with 16.83%and 37.56%.
Keywords/Search Tags:multi-agent reinforcement learning, robotic arm assembly, multi-agent system
PDF Full Text Request
Related items