Learning Grasping Skills Using Deep Reinforcement Learning

It is so difficult to establish a suitable mathematical model for the programmed robot and computer system under the non-structural environment with a large number of unknown objects that we can solve this problem by traditional methods.Andthis paper try to learn grasping skills using deep reinforcement learning method.Firstly,according to the characteristics of the crawling task,the problem of low sample utilization in strategy optimization for the deep strategic gradient(DDPG)algorithm is improved.The DDPG algorithm is improved,and the weighted sampling DDPG algorithm is proposed.The comparison experiment is carried out.The weighted sampling DDPG algorithm is two times faster than the DDPG algorithm.For the low-intensity learning process,the fusion algorithm of target detection and deep reinforcement learning is proposed.The region where the target object is located is a priori of the deep reinforcement learning algorithm,so that deep reinforcement learning can obtain more effective learning samples in the exploration phase,thereby improving the learning speed of the deep reinforcement learning algorithm;and guiding the method and reward in the simulation environment.The algorithm performs comparative experiments.The experiment shows that the learning speed of the algorithm is about 3 times higher than that of the guided algorithm.Secondly,for the scenes where the objects are densely distributed during the capture,the objects affect each other and the trajectory cannot directly sample the single object capture method.A dense Q network-based dense object capture sequence planning method is proposed.Easy-grabbing objects,clearly grasping obstacles for difficult-to-grab objects,and thus improving the success rate of capture.Simulation experiments verify that the method can improve the catching success rate by 35% compared with the unplanned method;Under the limitation of the scene,a grasping strategy of pushing and collaborating is proposed.When the object is densely distributed,the robot is used to push the object to create a larger grab space and improve the success rate.The simulation results show that the method can be better.It is difficult to capture a successful scene by simply grabbing the action.Finally,the performance of the method based on the deep reinforcement learning robot is evaluated in the simulation environment.The success rate of the method is stable at 85% in the simulation environment.At the same time,in order to directly use the deep reinforcement learning algorithm to train the actual robot,the robot loss is relatively large.The robot is pre-trained in the simulation environment,and the migration learning technique is added in the pre-training,so that the training strategy is closer to the real one.The environment has verified through experiments that the model parameters obtained by this migration method can be applied to the actual environment.