The Research On Grasping Path Planning For Moving Objects Based On LSTM-FC And Deep Reinforcement Learning

Highly intelligent manipulator grasping technology has always been one of the important goals of robotic research and development.The method of grasping randomly moving objects is an important function to realize the industrial assembly line from automation to intelligence.This paper studies two aspects of the problem of grasping moving objects by manipulators.Firstly,the traditional method of capturing moving objects based on the prediction mechanism has poor prediction accuracy.This paper proposes a prediction network that combines a long and short-term memory network model(LSTM)and a fully connected network to predict the movement trajectory of objects and improved trajectory prediction accuracy.Secondly,the method of grasping moving objects based on the prediction mechanism requires an accurate model of the manipulator.The poor real-time performance caused by the long prediction time makes it difficult to deploy to a real robot.In response to the above problems,this article focuses on a moving object grasping system based on the reinforcement learning algorithm Soft-Actor Critic(SAC),and the deployment of the system to real robots successfully achieved grasping of moving objects.The main work is as follows:(1)A moving object trajectory prediction model(LSTM-FC)combining long short term memory(LSTM)network and fully connected layer neural network is proposed.The model can predict the future movement trajectory of the object based on the movement characteristics of the historical trajectory of the object.This paper collects data based on Coppelia Sim as a simulation environment,uses multi-feature input to describe the movement information of the historical trajectory of the moving object,and then trains the network model to form the desired input-output mapping relationship,lastly predicts the future trajectory of the object.At the same time,a grasping simulation environment was built in Coppelia Sim and the environment used the prediction model to carry out a simulation experiment of Baxter manipulator for grasping moving objects.The LSTM-FC network model can not only process sequence data more accurately and effectively,but also has higher accuracy in predicting the movement trajectory of the object,thereby improving the grasping success rate.(2)The moving object grasping method based on the prediction mechanism requires an accurate model of the manipulator and the poor real-time performance caused by the long prediction time makes it difficult to deploy to a real robot.In response to the above problems,this paper designs a moving object grasping system based on the deep reinforcement learning algorithm Soft-Actor Critic(SAC).The system is divided into two parts:the object detection module and the manipulator control module based on deep reinforcement learning(DRL).The paper elaborated the detailed design of the two modules respectively.The manipulator control module adopts an integrated tracking+grasping strategy and consider keeping the relative distance between the end of the manipulator and the moving object stable as the goal to set the reward function.The system realized controlling the end of the manipulator to track moving objects steadily first,and then the integrated operation of grasping is performed.In addition,some optimizations have been made to the reinforcement learning algorithm SAC used to speed up the convergence.₂L regularization correction is introduced in the training process to avoid over-fitting during neural network training and improve the generalization of the deep reinforcement learning control strategy.The object detection module outputs the three-dimensional coordinates of the moving object in the manipulator base coordinate system in real time,so that the DRL-based control module is successfully deployed on the real robot.Finally,it is combined with the ROS-based robot control framework to perform multiple real robot grasps.In the experiment,the objects with different moving trajectories were successfully grasped.On the other hand,This paper test the grasping system for multiple times and record different grasping steps for manipulator.The data shows the feasibility and generalization of the grasping system.