Autonomous navigation of agricultural machinery has been the fundamental and core pillar in the advancement of precision agriculture. The production efficiency and product quality can be enhanced to simultaneously reduce the resource consumption and environmental footprint. This review aims to systematically examine the research progress in the autonomous navigation technology for agricultural machinery. A particular emphasis was placed on the current state-of-the-art and emerging trends within two critical technological domains: navigation positioning and navigation control. Significant efforts were also made on the navigation positioning. Specifically the high-precision global navigation satellite system (GNSS) (real-time kinematic (RTK) and precise point positioning (PPP)) was delivered the centimeter-level absolution of the positioning accuracy. Concurrently, machine vision approaches provided robust solutions for the centimeter-level relative positioning. Both conventional algorithms and increasingly powerful deep learning techniques were balanced for the feature extraction and scene understanding, along with light detection and ranging (LiDAR) systems. Some limitations were also proposed after the comparison. Single-sensor systems were recognized in the complex, dynamic agricultural environments. The dominant and most effective paradigm emerged as the multi-sensor fusion strategies integrating GNSS, inertial navigation systems (INS), vision, and LiDAR data. This convergence of data streams was essential to enhance system resilience, reliability, and adaptability across diverse and challenging field conditions. In navigation control technologies, solutions were tailored to different machinery classes. Hydraulic steering systems remained the preferred choice for heavy-duty agricultural equipment, due to their high force. While the electric steering systems offered distinct advantages, in terms of precision, responsiveness, and integration ease, increasingly suitable for medium and small-sized machinery. Substantial progress was achieved in the path-tracking algorithms. Some techniques, such as Model Predictive Control (MPC), were used to predict the future states for the optimal control actions. Adaptive Pure Pursuit methods dynamically adjust look-ahead distances for smoother tracking. Intelligent optimization algorithms further enhanced the accuracy and the robustness of path following. Advanced control strategies demonstrated the superior performance, particularly under demanding operating scenarios like steep slopes, uneven terrain, and conditions prone to wheel slip or side-slip. Technological evolution was found in the practical implementation and commercialization. International manufacturers of agricultural machinery, including John Deere and CLAAS, successfully transitioned the high-precision autonomous navigation solutions into the large-scale, commercially viable products widely adopted in modern farming. National enterprises, such as Huace Navigation and Shanghai Lianshi, also deployed navigation systems. BeiDou navigation satellite system (BDS) was also utilized under various agricultural scenarios. Looking ahead, future research should strategically focus on several key directions: The intelligent context-aware fusion for the multi-modal sensor data streams; navigation algorithms optimization specifically for highly specialized agricultural tasks and environmental conditions; and the deeper, more seamless integration of autonomous navigation with the precision agricultural implements and operations. These avenues can be instrumental for agricultural machinery autonomous navigation towards the higher levels of intelligence, broader applicability, and greater practical utility. This trajectory can provide a strong reference to accelerate the modernization and sustainable transformation of global agriculture.

Ensuring the accurate detection of pineapple fruits under the high planting density and serious homogenization represents a current and significant challenge. In this study, an enhanced lightweight detection framework, derived from the improved You Only Look Once version 5s (YOLOv5sp), is investigated in terms of the rapid and precise recognition of pineapple fruit for the agricultural robot. Three Convolutional Block Attention Module (CBAM) attention modules are considered the backbone network responsible for feature extraction, and the SIoU loss function is introduced to replace the CIoU loss function to handle the orientation angle and the penalization index. Eventually, the designed YOLOv5sp detection result of the mAP@0.5 value is 94.5%, which is 6.30% higher than YOLOv4, 1.83% higher than Faster R-CNN, and 6.90% higher than classical YOLOv5s. At the same time, compared with the models SHFP-YOLO and RGDP-YOLOv7-tiny in other pineapple detection literature, the mAP@0.5 of the designed model is 4.54% and 3.5% higher, respectively. Furthermore, when it comes to the agricultural robot operating in diverse natural situations, the YOLOv5sp algorithm can maintain a successful picking rate of 90% with an average time of 15 s, exhibiting the effectiveness of the visual component in engineering scenarios. These research results can accelerate the transition of pineapple harvesting from manual to automated operations.

Grape-picking robots can be an effective solution to deal with the contradiction between manual labor efficiency and the limited harvesting period, with the rapid development of machine vision and artificial intelligence. The varying sizes and shapes of grape key structures have limited the working space of the robot at the grape harvesting stage. Improper positioning of picking points can also lead to collisions between the robot's end and the grapes, even the damage and dropping. In addition, it is necessary to consider such collisions between the robot's end and the fruit branches. The reason is that these collisions can result in failed picking, damage to the branches, and the risk of fungal infection in the fruit trees. In this study, the localization algorithm was proposed for the picking points of grape key structures using deep learning and multi-object recognition. Picking point localization was enhanced to reduce the grape damage and the failure rate during harvesting. Firstly, the G-YOLACT++ model incorporated the SimAM attention module and Mish activation function to optimize the YOLACT++ model. Then the key grape structures were detected, such as grape-bearing branches, grape peduncles, and grape clusters. As such, these grape structures in the multi-adjacent clusters were segmented into multiple masks within the field of view. The membership of grape key structures was determined within the same cluster using their intersection and relative positions. The same string of grapes was then merged to select the Region of Interest (ROI) area with the low collision for grape pedicles. The range of re-selection was also designed to locate the picking point. The experimental results demonstrated that the incorporation of the SimAM attention mechanism into the YOLACT++ model resulted in an improved mean average precision (mAP) for the mask. The Mish activation function was selected to replace the ReLU in the backbone network. After that, the mAP values of the mask and bounding box increased by 0.3 and 2.23 percentage points, respectively. Both modifications were greatly contributed to the enhancement of the performance. The average mAP values of the bounding box and mask in G-YOLACT++ were improved by 0.83 and 0.88 percentage points, respectively, compared with the YOLACT++. By contrast, the mAP values of the improved model for the bounding box and mask increased by 2.36 and 2.13 percentage points, respectively, compared with the original. Furthermore, the sizes of all the improved models remained unchanged, while there was a relatively slight improvement in the inference speed. Therefore, there was a positive effect of improvement on the performance of the models. The correctness rates of the single and multiple fruit samples were 88% and 90%, respectively, for the key structure-dependent judgment and fusion. The correctness rate was 92.3% for the removal of grape clusters with the incomplete recognition of key structures. Compared to the two positioning methods that use the center of the bounding rectangle enclosing the grape peduncles in ROI and the centroid of the grape peduncles identified by the model as the picking points, the success rates of the picking point localization method in this study were improved by 10.95 and 81.75 percentage points, respectively. These results demonstrated the research could be a viable support to the optimization of grape picking robots and lays the foundation for low-damage harvesting of clustered fruits in unstructured environments.

To improve the GNSS receiver’s accuracy, continuity, and stability in measuring the height of agricultural implements, this study proposed a variable-parameter Kalman filter (VPKF) algorithm based on GNSS and accelerometer to estimate the height of the implements optimally. The VPKF was verified, and its accuracy was evaluated by parallel rail platform and field tests. From the parallel rail test results, when the GNSS receiver was in real-time kinematic (RTK) positioning and the time delay of differential correction data (TDDCD) was less than or equal to 4 s, the root mean square error (RMSE) of the VPKF estimation was 9.82 mm. The RMSE of the GNSS measurement was 18.85 mm. When the GNSS receiver lost differential correction data within 28 s, the absolute error of VPKF was less than 30 mm, and the RMSE was 16.93 mm. The field test results showed that when the GNSS receiver was in RTK positioning and the TDDCD was less than or equal to 4 s, the RMSE of VPKF estimation was 13.43 mm, and the GNSS measurement was 14.56 mm. When the GNSS receiver lost differential correction data within 28 s, the RMSE of the VPKF estimate was 15.22 mm. These results show that VPKF can optimally estimate implement height with better accuracy. Overall, the VPKF can obtain a more accurate, continuous, and stable height of the implement, and increase the application scenarios of the GNSS receiver to measure the implement height.

The primary aim of this study was to classify the hazard level of brown planthopper (BPH) damage in rice. Three datasets, including spectral reflectance corresponding to the sensitive wavelengths from rice canopy spectral wavelengths, rice stem spectral wavelengths, and fusion information of rice canopy and stem spectral wavelengths were used for BPH hazard level classification by using different algorithms. Datasets and algorithms were optimized by the BPH hazard level classification effects (which was evaluated by indices of accuracy, precision, recall, F₁, and k-value). The optimized algorithm combination was used to build a hazard level classification model for spectral reflectance corresponding to the sensitive wavelength from the rice canopy spectral images. Results showed that: (1) The spectral reflectance corresponding to the sensitive wavelengths of fusion information dataset performed best in BPH hazard level classification, with the highest accuracy (99.08%), precision (99.31%), recall (98.83%), F₁ (0.99), and k-value (0.99). (2) The optimum algorithm combination was Savitzky-Golay (S-G) smoothing, principal component analysis (PCA) for sensitive wavelength selection, and broad-learning system (BLS) for modeling. (3) The spectral reflectance corresponding to the sensitive wavelengths dataset of rice canopy spectral images achieved accuracy (80.63%), precision (80.28%), recall (77.03%), F₁ (0.79), and k-value (0.74) in classifying BPH hazard level by using the optimum algorithm combination.

"Who will farm" and "how to farm" would be one of the common problems in agriculture all over the world in the future. Modern agriculture is the development direction for agriculture, and it is an effective way to solve the problem of "who will farm" and "how to farm". Smart agriculture has been the promising development direction in modern agriculture. Among them, the unmanned farm is one of the most important ways to realize smart agriculture. South China Agricultural University first constructed the unmanned rice farm in Zengcheng, Guangzhou, Guangdong Province of China in 2020. This unmanned rice farm has broken through four key technologies, including (1) Digital perception for the accurate perception of the operating environment, working objects, and machinery information of unmanned farms; (2) Intelligent decision-making for land reclamation, tillage, planting, sowing, field management, and harvest programs; (3) Precision operation for the automatic navigation and precision of agricultural machinery; (4) Intelligent management for crop growth, agricultural machinery operation, and farm management. Five functions were achieved, including (1) Full coverage production links of tillage, planting, field management, and harvesting; (2) Automatically transferring from the garage to the field; (3) Automatic obstacle avoidance and parking for safety; (4) Real-time monitoring of crop production, and (5) Intelligent operation with the decision-making precision. Remarkable economic, social, and ecological benefits were also achieved in this unmanned rice farm. The yield of the Simiao rice variety was 9934.35 kg/hm² in the unmanned rice farm in Zengcheng, Guangdong Province of China in 2021, 32% more than the local average yield of the same rice variety. The yield of ratoon rice was 17 988 kg/hm² in an unmanned farm in Qianshanhong Town of Yiyang, Hunan Province of China in two seasons of 2023. Therefore, the crops were planted well in the unmanned farm. 30 unmanned farms were constructed in 14 provinces in China by the end of 2022, including rice, wheat, corn, and peanuts unmanned form. The unmanned farm can provide the great potential to solve the "who will farm" and "how to farm" in the future. In order to further promote the development of unmanned farms, the following recommendation were proposed: (1) Enlarge the agricultural farmland scale and agricultural management scale, strongly support the agricultural cooperative, agricultural leading enterprises and large farm landowners to enlarge their farmland scale to suited the construction of the unmanned farm. (2) In light of the basic requirements of unmanned farm for lager farmland, flat field, continuous planting, and convenient farm track path and irrigation and drainage channel, land improvement and develop high-standard farmland should be strengthen . (3) Explore suitable unmanned farm models for different area, different agrotype, different crops, different cropping systems, and different production models. It is necessary to promote the construction of unmanned farms according to the local modern agricultural industrial parks. (4) Explore various talent training for unmanned farm. For example, it would be a good propose to train technicist and manager working for unmanned farm in vocational school. (5) It is necessary to speed up the formulation of relevant policies to promote the construction of unmanned farms, especially the financing of construction funds. To mobilize the enthusiasm of the government, enterprises and society to invest in the construction of unmanned farms, and appropriately increase the purchase subsidy amount of intelligent agricultural machinery should also be considered as soon as possible.