Research On Grasp Control Of Robotic Arm Based On Depth Image And Deep Reinforcement Learning

In order to enable robotic arm to grasp objects with unknown geometric structure and to plan the motion path autonomously when the position of the object changes,this paper uses image depth information and convolutional neural network to realize the robotic arm to predict the optimal grasping pose,and uses deep reinforcement learning algorithm to achieve robotic arm performs the grasping operation autonomously.For the task of predicting the best grasp pose,this paper improves the generative grasp convolution neural network in network depth,convolution kernel size and other aspects.Training and testing the improved neural network using samples from Jacquard grasp detection dataset.Samples from Jacquard dataset are extracted to form a validation set to verify the network generalization performance,achieving the highest prediction accuracy of 87.5%.The six-degree-of-freedom simulation manipulator equipped with a depth camera is equipped with a network model with good generalization performance,and experiments of grasping objects are carried out in the robot operating system.For static objects,open-loop grasping is used,and for dynamically moving objects,visual servoing is used for closed-loop controlled grasping.Each experiment finally achieved more than 80% grasping accuracy on the test set.For the task of autonomously controlling the end of the robotic arm to reach the target area,a two-dimensional simulated robotic arm is constructed,and the deep deterministic policy gradient algorithm is used to train the robotic arm to complete the task autonomously.The algorithm system is improved in terms of reward function and replay experience sampling strategy,and compared through simulation experiments.The experimental results show that both improved methods can improve the convergence speed and stability of the robotic arm when performing tasks.In order to realize the robot arm grasping the objects autonomously,a three-dimensional simulation robot arm is built in the robot operating system.The improved scheme of the deep reinforcement learning algorithm was transferred to it.Experimental comparison showed that the improved scheme was beneficial to the deep reinforcement learning algorithm to train the robotic arm to improve the success rate of grasping objects.