基于深度强化学习的仿真机器人轴孔装配研究

Deep Reinforcement Learning Based Robotic Assembly in Simulation

机器人轴孔装配问题是机器人装配任务中的关键问题,制造业中大量的操作任务可以归结为轴孔装配问题。传统的机器人轴孔装配方法往往依赖固定的结构化操作环境,且仅限于特定几何形状产品的装配,操作环境的变化很容易导致装配任务失败,会损害产品部件或者导致机器人末端受到很大的力和力矩而损失机器人精度。随着智能手机等行业个性化定制的需求的不断提出,生产线往往需要快速变更,而传统的机器人任务往往依赖大量的手动编程和硬编码的方法,难以适应这种快速的产品线变更。基于传统方法的这种局限性,提出在3D图形仿真环境下基于深度强化学习的自学习方法来训练机器人学习轴孔装配任务,利用深度神经网络来编码机器人控制策略,通过强化学习方法来训练该深度神经网络,以泛化机器人的操作策略,使得机器人能处理更多环境不确定性因素下的轴孔装配问题。上述方法不需要进行大量的人工编程和硬编码策略。最后通过基于ROS和Gazebo的仿真环境验证了仿真环境下使用上述方法进行轴孔装配仿真的有效性。

Robotic peg-in-hole task is a key problem in robotic assembly tasks,and many robotic assembly manipulation tasks can be treated as the peg-in-hole tasks in manufacturing industry.The traditional peg-in-hole methods depend on the fixed and structured environment,and can always be applied in some specific geometric products,which may cause the task failed if the environment was changed and destroy the products or lose the robot precision when the robot end-effector is subjected to a great force.In the smart phone manufacturing industry,products need personalized customizations,and this requires frequent changes in the production line.The traditional robot tasks often rely on a large number of manual programming and hard coding methods,and it is difficult to adapt to this rapid product line changes.In view of the limitations of the traditional method,this paper proposes a self-learning method of robotic peg-in-hole task based on deep reinforcement learning.This method uses deep neural network as the control policy of robot,and uses reinforcement learning method to train the deep neural network to improve the generalization ability of the robot control policy which can makes robot adapt to more uncertainty.This method does not require a lot of manual programming and hard coding strategy.The article also verified the validity of the method in a simulation environment based on ROS and Gazebo.