Target Grabbing Control Of Underwater Robot Based On Deep Reinforcement Learning

In order to protect the marine environment and maintain sustainable development while acquiring marine resources,China has built a large-scale marine pasture.Today,China has become a large aquatic product country.However,some aquatic organisms such as sea cucumbers and sea urchins require divers to go fishing.This work is difficult and dangerous,so the cost is high,which has led to the idea of using underwater robots(Remote Operated Vehicle,ROV)instead of divers to go fishing.Because most of the current ROV autonomous grabbing methods have some defects,which cause the reason of low efficiency,they cannot be put into production.Therefore,this paper hopes that deep reinforcement learning can be used to solve the task of target grabbing by underwater robots and verified in a simulation environment similar to the real situation.The main research contents of the paper are as follows:According to the entity of ROV,model it in the mujoco software,and simulate the pool environment to set the specific density and viscosity in the environment.Then,the kinematics and dynamics of the ROV are analyzed,including how the density and viscous forces on the ROV when it moves in the environment are calculated,and the simulation environment is simulated from simple to difficult.The algorithm provides a platform for simulation verification,and the animation can be used to show the ROV grabbing process.Analyze the thrust distribution of the underwater robot and the influence of the gripping process of the robotic arm on the trim.Set the appropriate gripping attitude during gripping so that it can complete two different gripping methods,remote gripping and full autonomy.Analyze the shortcomings of the two grabbing methods.The remote capture method is used to collect the data needed for supervised deep reinforcement learning.Research on the use of supervised deep reinforcement learning to complete the capture task,hoping to use the characteristics of long and short memory networks and mixed density networks to train the data of human-operated underwater robots to complete the capture task,so that the network after training can be targeted If the information is not completely correct,the crawl task is finally completed.Due to the difficulty of sampling,the research was performed only in a simulation environmentIn order to further improve autonomy,ROV can complete target capture efficiently in a simulation environment closer to the real environment.Research on deep reinforcement learning methods.In order to avoid the situation of sparse reward,the reward function is set.Analyze the effect of the algorithm after training at the same time,improve the best-effort near-end optimization strategy,use KL divergence instead of the standard deviation term in the diagonal Gaussian strategy,so that it can effectively avoid convergence to the local optimum excellent.And it is gradually trained in the simulation closer to the real environment,and achieved good results.