Robot Push And Grasp Manipulation Skills Learning Based On Deep Reinforcement Learning

Push and grasp manipulation skills are the basis for robots to handle object grasping tasks.The existing push and grasp manipulation skills learning methods didn't consider the shape of objects and prior knowledge of environment when grasping a object,and can't function well when multiple objects are mixed.This paper studies push and grasp manipulation skills learning based on deep reinforcement learning.Two sets of push and grasp manipulation skills learning methods are proposed for single-object grasping tasks and multi-object grasping tasks.Based on RGB-D sensor and UR5 manipulator,a robotic push and grasp manipulation skills learning system using deep reinforcement learning is built to verify the effectiveness of the proposed methods.Aiming at the problem that the existing methods didn't consider the object shape and prior knowledge of the environment when grasping a single object,the push and grasp manipulation skills learning method that considers object shape and prior knowledge of the environment is studied.In order to employ prior knowledge of object shape and environment to formulate targeted manipulation strategies,a DQN decision model with SVM reward module is proposed for learning multi-step push or grasp manipulation.The key of the proposed model is that the prior knowledge of object shape and environmental constraints is introduced into the reward system by a SVM classification model.Therefore,the reward system will continuously motivate the robot to approach to the target manipulation during training process.The experimental results in the actual environment show that the method reaches a success rate of90% in grasping cube objects,and the success rate of pushing cylindrical objects to the constrained boundary has reached 75%.Aiming at the difficulty of grasping when the multi-objects are mixed,an improved push and grasp sequence manipulation skills learning method for multi-object is studied.In order to prevent the multi-object grasping method combined with push and grasp manipulation skills from excuting grasp manipulation when it is not suitable for grasping,a grasping judgment module is introduced to prejudge whether the situation is suitable for grasping.Considering that the push manipulation sometimes only pushes multiple objects as a whole instead of pushing them apart,the reward setting improvement is implemented to motivate the DQN decision model to learn push manipulation skills that will push away multiple objects and not push them outside the workspace.The improved method effectively improves the efficiency of multiobject grasping.According to above research on robot push and grasp manipulation skills learning using deep reinforcement learning method,a robotic push and grasp manipulation skills learning system based on deep reinforcement learning was built,and three modules of feature extraction,motion decision-making and motion planning were developed,which verified the feasibility and effectiveness of the method proposed in this paper.