Research On Picking Arm Grasp Method Combining Attention Detection And Reinforcement Learning Viewpoint Planning

With the development of high-tech with computers as the core,the cost of robots has significantly decreased,and using picking arms to harvest fruits is the future development trend.However,factors such as changes in lighting,obstruction of branches and leaves,and fruit overlap have always been important factors affecting harvesting efficiency.Aiming at the problem of low accuracy in detecting fruit targets with changes in lighting,overlap,and occlusion in natural environments,a method for anchor free fruit detection and segmentation based on dual attention guided networks is proposed,and the effectiveness of the algorithm is verified through experiments.At the same time,in order to solve the problem of low grasping accuracy of the picking arm in environments with overlapping fruits and obstructed branches and leaves,a deep reinforcement learning based picking arm viewpoint planning grasping method is proposed,which enables the picking arm to grasp fruit targets more efficiently and accurately.Simultaneously integrate a harvesting robot system to verify the effectiveness of the method.The main work of the thesis is as follows:(1)Propose a fruit object detection and segmentation method based on dual attention guided networks to improve the effectiveness of fruit detection in complex background environments.Firstly,in order to address the challenge of lighting changes,an image correction module is introduced to adaptively convert images under different lighting conditions into similar lighting,improving the network’s robustness to natural environmental lighting changes.Secondly,in order to enhance the ability of network feature extraction,a multi-scale enhanced fusion feature pyramid network is adopted as the feature extraction module,effectively reducing the interference of complex backgrounds on fruit detection results without increasing computational complexity.Finally,due to the inability to accurately identify the boundaries of occluded and overlapping fruits,a dual attention guided mask branch is introduced to focus on the pixels of irregularly occluded and overlapping objects,thereby further improving detection accuracy.The experimental results indicate that this method can reduce the impact of complex natural environments,thereby effectively improving the effectiveness of fruit detection.(2)Propose a deep reinforcement learning based picking arm viewpoint planning and grasping method to improve the accuracy of fruit grasping in complex occlusion environments.This method is based on the DDPG(Deep Deterministic Policy Gradient)algorithm and introduces GRU(Gate Recurrent Unit)into the DDPG network.It utilizes GRU’s memory ability to optimize the DDPG network structure,improve network stability,and accelerate algorithm convergence.At the same time,a mixed noise composed of Gaussian noise and Ornstein Ulenbeck noise is designed to enhance the randomness of the learning process of the picking arm,avoid falling into local optima,and improve learning efficiency.Finally,optimize the reward function to guide the picking arm to grab towards the target point more stably.The experimental results show that this method can accelerate the convergence speed of DDPG,improve the efficiency and success rate of viewpoint planning and grasping.(3)Based on the core algorithm proposed above,integrate a harvesting robot system in a simulated environment.The system is composed of a six degrees of freedom picking arm,three finger flexible gripper,mobile chassis and other hardware.At the same time,a set of visual control software is developed and realized.Based on this system,the effectiveness of the proposed object detection method and reinforcement learning viewpoint planning grasping method is verified in a simulated environment,which helps to promote the industrialization and practicality of domestic fruit picking robots.