Research Of Path Planning For Greenhouse Weeding Robotic Arms Based On Deep Reinforcement Learning

The trend in greenhouse crop seedling stage is towards the use of intelligent and mechanized weed management technology.Researchers are increasingly focusing on the development of intelligent mechanical weeding robots.However,there are still some challenges in this field,such as low precision of crop identification through visual detection,limited weed species identification,and difficulty in achieving active fixed-point weeding research while avoiding damage to seedlings.To address these challenges,a path planning algorithm for the weeding robot arm based on an improved Deep Deterministic Policy Gradient(DDPG)has been proposed.An improved Faster R-CNN algorithm for target detection and location of seedlings and grasses was proposed to achieve the visual location of plants required by DDPG algorithm training.The DDPG algorithm is improved by introducing the rules of reward equipotential surface.A dynamic weeding test prototype was designed,and the performance of the model was tested in a greenhouse.The verification results show that the proposed algorithm can meet the requirements of actual environment weeding operation.The main conclusions obtained are as follows:(1)Research on seedling and grass target detection and localization based on deep learning.A crop and weed dataset with a total of 23,000 images was developed.By comparing the performance of Faster R-CNN model of Res Net50,Res Net101,VGG-16 and VGG-19 backbone feature extraction networks,VGG-16 network was determined as the optimal feature extraction network.On this basis,combined with attention module,super green index and maximum threshold variance method,Faster R-CNN model was improved.It has been proved that compared with the model before improvement,the detection accuracy is improved,the detection time is shortened,the model can accurately identify lettuce seedlings,and the green plants and background other than the seedlings can be divided and marked as weeds,and the positioning coordinates can be output at the same time,which provides the basis for the path planning research of the weeding robot.(2)Path planning of the weeding robot arm based on deep reinforcement learning.The principle of depth deterministic strategy gradient algorithm(DDPG)is described.DDPG network is trained in Coppelia Sim software and converges successfully.In order to analyze the shortcomings of the algorithm,an improved DDPG algorithm with artificial potential field method is proposed.The corrected algorithm was 93.36%,2.79%,and 57.73 ms for each weeding path.Compared with previous studies,the model has excellent comprehensive performance and can meet the requirements of the weeding robot.(3)Design of a dynamic weeding test platform.The actual greenhouse environment was investigated,and the overall working plan of the dynamic weeding test platform was analyzed and designed.The mobile platform is designed,and the type selection and parameters of the manipulator are designed.A vertical weeding end-effector was designed.By finite element analysis,it is found that the suitable burial Angle is 60°.Based on the above,the design of dynamic weeding test platform is completed,which provides a hardware basis for the performance verification test of the weeding robot path planning algorithm in the actual environment.(4)Performance Testing and Weeding Experiment of the Prototype: The design of the dynamic weeding platform has been completed,and the Orbbec Astra Pro binocular camera height measurement test was completed.The central control software was written in Python language.The herbicidal performance test was conducted in an actual greenhouse environment.At a speed of 0.10m/s,the average detection accuracy was 98.09%,the average weed segmentation accuracy was 93.41%,the average weeding success rate was 91.00%,and the average seedling injury rate was 3.80% in a single crop row.These experimental results indicate that the proposed algorithm,which is based on the improved Faster R-CNN seedling and grass target detection and the improved path planning algorithm based on DDPG weeding robot,can successfully meet the requirements of an active and accurate weeding operation between plants.