The Target Pose Perception And Grasp Skill Learning Orient To Manipulator Autonomous Operation

Using manipulator to perform operation tasks intelligently and efficiently has become a hotspot in the field of robotics research.However,because of the complex and variable application scenarios,robots need to adopt a more intelligent control method to complete operation tasks more autonomously in unstructured environments.In view of the many difficulties in autonomous operation,this paper provides a variety of solutions based on intelligent learning methods,using deep learning to extract environmental characterization information from vision,accurately sensing the target posture state,and using visual servo control to complete the operation.In addition,deep reinforcement learning is used to allow robots to interact with the environment to learn the optimal control strategy automatically,and to optimize problems such as sparse rewards.This paper studies from the following directions:First,the Kinova 7DOF redundant manipulator forward and inverse kinematics solution and optimization problems were studied,for target recognition and pose estimation tasks,Label Fusion was used to produce a data set for supervised deep neural network training,the manipulator and operation were written The physical model is built by Mujoco to build a physical simulation environment for training robot agents to learn operation skills autonomously.Secondly,for the autonomous operation task in complex unknown environment,a feature fusion algorithm based on deep learning network is proposed.The color image output from the RGB-D camera is used to identify the target of interest and segment it from the background,and then based on the recognition segmentation result To achieve pixel-by-pixel fusion of RGB and Depth features,the global feature provides richer and more accurate details for the posture prediction module,the output posture is further fine-tuned by point cloud registration,and finally the accurate target operation posture is obtained.This algorithm can effectively overcome the influence of environmental factors such as occlusion,shadow,light change and so on,and has strong robustness.However,due to the complex network and post-processing of the results,this solution cannot meet some of the high-speed operation tasks,so this paper additionally designed an end-to-end network model based on CNN to estimate the target grasping pose,using only depth cameras Observing the target can generate the optimal grab posture in real time.Thirdly,the deep reinforcement learning framework is used to allow robots to learn sequential decision-making skills such as grasping and handling.A multi-module deep neural network is designed to extract features from operating tasks.A stacked LSTM network is proposed for time series learning.A distributed PPO algorithm is used to implement agents and The environment interactively updates the network parameters,and then obtains an end-to-end visual drive strategy from the camera-side image input to the robot arm joint state.In addition,a small amount of demo sample data is added to the memory playback unit to initialize agent parameters and guide strategy search to improve training efficiency.In order to ensure the safety of the learning process,the model is trained in the Gym and Mujoco environments using the random domain method,so that the policy parameter network is exposed to a large number of random simulations to fully express the physical characteristics of the real environment,so that the driving strategy for simulation training can be directly Transplanted to the actual robot.Finally,a robotic arm visual servo control platform was built based on ROS to verify target perception effects and collect operation demonstration videos.In addition,for the above studies,physical or simulation experiments were designed to verify the feasibility,and the results and algorithm performance were evaluated and optimized.