A key challenge for automated orchard management robots is the rapid and accurate identification of crop growth and maturity conditions for subsequent operations, such as automatic pollination, fertilization, and picking. In particular, strawberries have a short ripening period and the fruits are heavily overlapped and shaded by each other, which is time-consuming and inefficient under traditional detection methods. Therefore, we designed and developed a strawberry growth detection algorithm, SDNet (Strawberry Detect Net). The algorithm is based on the YOLOX model and replaces the original CSP block in the backbone network with a self-designed feature extraction module C3HB block to improve the spatial interaction capability and monitoring accuracy of the detection algorithm; Then, the normalized attention module (NAM) is embedded in the neck to improve the detection accuracy and attention weight of small target fruits; and we use the latest SIOU objective loss function to improve the prediction accuracy of the detection model, which finally achieves the monitoring of strawberry fruits under five growth states. The experimental results show that the mAP, precision, and recall of SDNet are 94.26%, 93.15%, and 90.72%, respectively, and the monitoring speed is 30.5 ms. It is 4.08%, 3.64 and 2.04% higher than the precision, accuracy, and recall of YOLOX, respectively, and there is no significant change in the model size. The research results can effectively solve the problem of low accuracy of strawberry fruit growth state monitoring under complex environments, and provide important technical reference for realizing unmanned farm and precision agriculture.

The fruit and vegetable classification problem is an inseparable branch in the field of image recognition. GoogLeNet provides a more ideal solution for the fruit and vegetable classification problem. We use the GoogLeNet network to classify apples, lemons, oranges, pomegranates, tomatoes, and colored peppers. Through experiments, we obtained the training accuracy of GoogLeNet as 96.88%, the testing accuracy as 96%, and the training speed as 11.38 sheets/second. The recognition accuracy of this model can meet the basic recognition requirements, but the training speed is low. Therefore, we decided to optimize GoogLeNet to significantly improve the training speed and further enhance the recognition accuracy of GoogLeNet. We reduced the number of convolutional kernels of GoogLeNet and adjusted the structure of Inception, which reduced the number of parameters of GoogLeNet by nearly 48% and increased the training speed of GoogLeNet from 11.38 to 33.68 sheets/second. To further improve its recognition accuracy, we tried two methods: 1) introducing a new activation function Swish; 2) between convolutional layers, we introduced DropBlock layer; these two methods improved the testing accuracy of GoogLeNet by 2%. Finally, we introduce AlexNet, VGGNet, ResNet18, DenseNet121, and Inception-ResNet to compare with our optimized GoogLeNet. By comparison, we found that our model has incredible performance in ACC, AUC, FPS, Recall, etc.

In response to the navigation requirements of robots in orchard environments, this paper presents a navigation method for orchard robots based on point cloud maps. First, point cloud maps are processed with pass-through filtering and PCA algorithms to make them suitable for path planning. Next, tree rows within the orchard are clustered and segmented based on map orientations. Then, the least squares method and the $\text{A}^{\ast} $ algorithm are combined for global path planning on local maps. Lastly, the TEB local path planning algorithm is employed to ensure that the robot navigates along the operation path. Experimental results indicate that the robot can successfully navigate orchards at speeds ranging from 0.4 to 1.0 m/s. The average longitudinal deviation obtained under these conditions is 26.7 cm, with a maximum value not exceeding 46.2 cm. The average heading angle deviation is 4.09°, with a maximum value of 8.65°. In conclusion, this approach ensures comprehensive navigation for orchard robots and lays a strong foundation for future research.

Accurate fruit tree models are essential for canopy volume measurement work, we build an orchard mobile robot platform and develop a fruit tree model reconstruction algorithm based on it, optimize the LIDAR Odometry specifically for the orchard environment, fuse the LIDAR Odometry, Inertial Measurement Unit (IMU), Global Navigation Satellite System (GNSS) sensor information and loop closure detection in the form of factors to add factor maps for back-end optimization to reconstruct the orchard map model, use the sliding window method to process in real time The fused information and narrowed the range where the tree trunks are located for two times of line surface feature matching, and the point cloud data are processed to get the fruit tree model. In order to make the point cloud distribution of the reconstructed model uniform, the robot also needs to match a specific walking route, and use the Hough transform and K-Means clustering algorithm to extract the linear-circular-linear walking route autonomously according to the tree row arrangement. The experimental results show that the error of the map model is less than 0.160 m, and the correlation coefficient R ² and root-mean-square error (RMSE) of the canopy model for volume measurement are 0.984 and 0.102 m ³ , respectively. The collected LiDAR data based on the robotic platform meets the requirement of fruit canopy volume calculation.

Existing target detection models are large and have multiple network parameters, which can severely slow down the detection speed when deployed on small, low-cost GPU-free Industrial Personal Computers (IPC). As a result, this study proposes a lightweight real-time tomato detection and point-picking integrated network model based on YOLOv5 (TDPPL-Net). Firstly, the algorithm replaces the YOLOv5 backbone with a four-group lightweight downsampling model consisting of Ghost Conv and Ghost Bottleneck to reduce the model size, while adding the attention mechanism SimAM module to improve detection accuracy after each scale’s feature map. Secondly, the Spatial Pyramid Pooling-Fast(SPPF) network structure is used and the convolutional layers in the (Feature Pyramid Network and Path Aggregation Network)FPN+PAN structure are replaced with a depth-separable convolution to reduce the computational effort. Finally, the center of the bounding box is used as the picking point, and the corresponding depth information is obtained in combination with the Intel RealSense D435 camera, which is converted into 3D coordinates under the robot arm coordinate system after hand-eye calibration. The experimental results show that TDPPL-Net reduces the number of parameters by 59.84% compared with the original YOLOv5, the model volume is only 40% of the original, the mAP is 93.36, and the real-time detection speed on the IPC without GPU acceleration is 31.41 FPS, which is 170.31% higher than YOLOv5. The TDPPL-Net increases detection speed on low-performance equipment without compromising detection accuracy. It can detect and locate tomato picking points in real-time in the complex natural environment, which can meet the working requirements of harvesting robots.

The investigation of the mechanical properties of rock mass can be effectively carried out through rock-like material experiments. In this study, polystyrene foam particles were utilized as a novel material for simulating initial damage within rocks. Our research involved the development of sandstone-like materials with comparable mechanical properties to actual sandstone. These materials were then subjected to orthogonal mechanical tests, allowing us to identify the key factors that have a substantial impact on the mechanical parameters of sandstone-like rocks. The influencing factors considered in the orthogonal mechanical tests were the proportion of aggregate and binder, the proportion of polystyrene foam in the entire model, the proportion of binder and regulator, and the size of polystyrene foam. Five levels were set for each factor, and mechanical parameters such as compressive strength, tensile strength, elastic modulus, axial strain, and Poisson’s ratio were tested for each group of samples. The changes in mechanical parameters with the levels of the above four factors were studied. The study found that modifying the proportion of aggregate to binder can alter the elastic modulus, tensile strength, and compressive strength values of sandstone-like material. The size of polystyrene foam can be modified to alter the axial strain values of sandstone-like materials. Additionally, adjusting the ratio of binder and regulator can modify the value of Poisson’s ratio. The comparison of mechanical parameters between sandstone-like samples and sandstone reveals that sandstone-like materials can better simulate the deformation and failure mechanisms of sandstone. The error in the main mechanical parameters, such as modulus of elasticity, strength, and Poisson’s ratio, is less than 7%, indicating a greater resemblance between sandstone-like materials and sandstone. Therefore, sandstone-like materials can be used to investigate the deformation law, damage evolution law, and failure mechanism of sandstone. This can help alleviate the difficulty of obtaining specimens of deep damaged rock and the high cost of testing.