In the agricultural domain, user demand texts serve as essential primary sources for agricultural extension, production management, and policy services. However, these texts typically contain highly specialized terminology, exhibit non-standard, colloquial, and diverse linguistic expressions, present fragmented semantics, and rely heavily on contextual reasoning. Such characteristics make them difficult to parse accurately using traditional rule-based approaches or shallow machine learning models. Consequently, these limitations often lead to biased demand classification and incomplete extraction of key factors, thereby constraining the quality of data available for intelligent agricultural decision-making. To address these challenges, the aim of this research is to develop a robust, domain-adapted, and highly interpretable structured analysis method for agricultural user demands.

Agri-NeedAgent, an agricultural user demand analysis framework, was proposed based on a "three-stage training + multi-agent collaboration" paradigm. First, during the domain knowledge pretraining stage, 80000 agriculture-related texts, including crop cultivation manuals, pest and disease control guides, agricultural policy documents, and farmer consultation records, were used to construct domain-specific semantic understanding, thereby enhancing the model's capability to interpret agricultural terminology, dialectal expressions, contextual logic, and implicit semantics. Second, in the instruction fine-tuning stage, 6320 annotated samples in an "instruction-input-output" format were employed to establish an explicit mapping from raw demand texts to structured outputs. Third, in the agricultural knowledge low-rank adaptation stage, Low-rank Adaptation（LoRA）was applied to perform lightweight parameter tuning on task-specific agents, enabling targeted adaptation for demand classification and key-factor extraction tasks. Built upon the above training process, a multiagent collaborative framework was constructed, in which the manager agent was responsible for task scheduling and quality control, while task agents were designed to perform demand classification, key-factor extraction, and explanation generation, respectively. Through this division of labor and collaborative mechanism, the framework achieved efficient and structured analysis of agricultural user demands.

Experimental results demonstrate that the proposed Agri-NeedAgent achieved a demand classification accuracy of 84.6%, a key-factor extraction F₁-Score of 85.2%, a structured interface compliance rate of 94.2%, and an interpretability score of 90.2.These results showed clear improvements over traditional deep learning models such as Bidirectional Encoder Representations from Transformers（BERT）as well as general-purpose large language models（LLMs）without domain adaptation. The findings confirmed the critical role of domain knowledge injection, explicit task alignment, and multi-agent specialization in enhancing semantic understanding and structured analysis of agricultural texts. Ablation experiments further validated the effectiveness of each component. Removing domain pretraining or LoRA fine-tuning resulted in substantial performance degradation in classification and key-factor extraction, indicating the necessity of domain adaptation and task-specific optimization for handling non-standard agricultural expressions. Moreover, eliminating the manager agent or the Reasoning and Acting（ReAct）mechanism significantly reduced structured interface compliance and interpretability, highlighting the importance of task coordination, intermediate verification, and multi-step reasoning for ensuring logical consistency and output completeness. Additionally, removing the external knowledge base reduced the interpretability score from 90.2 to 77.6, underscoring its essential role in providing theoretical grounding, reasoning support, and professional explanations. Although the multi-agent collaboration introduced an additional inference overhead of approximately 140 ms, the overall per-sample inference time remained within 225 ms, meeting the real-time requirements of agricultural consultation scenarios.

Supported by a "three-stage training + multi-agent collaboration" framework, LLMs can effectively address challenges posed by non-standard expressions, semantic fragmentation, and multi-factor reasoning in agricultural user demand texts. The proposed method demonstrated significant improvements in demand classification, key-factor extraction, structured output compliance, and interpretability, providing high-quality and traceable structured data for intelligent agricultural decision-making. After domain adaptation and task-specific tuning, the model not only gains enhanced capability for deep semantic analysis of agricultural user demands but also ensures the completeness and interpretability of outputs through multi-agent coordination. Although the current workflow still requires optimization in terms of data preparation, staged training, and knowledge-base updating, future work will focus on expanding region-specific and emerging-technology-related demand data, developing a dynamically updated agricultural knowledge system, improving multi-agent coordination efficiency, and exploring cross-lingual agricultural demand analysis to further promote the application and deployment of agricultural large models across broader scenarios.

Reinforcement learning (RL) has emerged as a core methodology in intelligent robotics, enabling autonomous agents to interact effectively with dynamic and uncertain environments. Despite its growing importance, RL instruction often remains disconnected from practical engineering applications, which limits students’ ability to translate theoretical knowledge into functional robotic manipulation systems. To address this challenge, this paper presents the design of an experimental scheme for robotic precise visual manipulation based on reinforcement learning and machine vision. The proposed scheme serves both as an instructional platform for deepening theoretical understanding and as a practical environment in which learners can engage with the complete workflow of robotic perception, decision-making, and execution in complex manipulation tasks.

A general-purpose robotic simulation platform was developed using CoppeliaSim as the core environment and integrated with a UR5 robotic manipulator and an RG2 gripper as the execution unit. An RGB-D camera was employed to acquire real-time workspace information, providing synchronized color images and per-pixel depth data, thereby substantially improving robustness in target detection and six-degree-of-freedom pose estimation. Within the perception module, RGB-D inputs were processed through height map generation, image preprocessing, and DenseNet-based feature extraction. The decision module employed a Deep Q-Network (DQN) to evaluate and select pushing and grasping actions. To explicitly connect theoretical instruction with hands-on experimentation, a dual-chain teaching framework was introduced. The theory chain focuses on control fundamentals, perception modeling, and decision optimization, while the practice chain emphasizes scene analysis, simulation-based data collection, and systematic parameter tuning. In addition, a Task Abstraction Layer (TAL) was implemented to facilitate rapid transfer of the learning framework to new manipulation tasks, such as flexible cable assembly, thereby demonstrating the platform’s scalability and generalization capability.

A series of simulation experiments were conducted to assess the effectiveness of the proposed experimental scheme. The results indicate that the robot successfully performed coordinated push-and-grasp operations in unstructured environments, even in the presence of significant object occlusion and scene clutter. Heatmap visualizations of learned Q-values reveal that the DQN-based decision module progressively refined its action selection strategy, converging toward optimal grasping behaviors. Parameter sensitivity analyses further show that both the discount factor and reward weighting exert a strong influence on training performance. A discount factor of 0.7 achieved the best balance between short-term and long-term rewards, yielding stable convergence and a grasping success rate exceeding 90% after 2000 training steps. Modifications to the reward function underscore the need to balance pushing and grasping incentives: eliminating pushing rewards slowed early-stage learning, whereas excessive weighting biased the policy toward pushing actions. Moreover, experiments using the TAL demonstrated that the framework could be adapted to a cable insertion task within approximately three hours, requiring fewer than 50 lines of code changes, thereby confirming its high reusability and adaptability.

The proposed experimental scheme effectively integrates reinforcement learning and machine vision within a hands-on educational framework for robotic manipulation. By emphasizing the explicit relationship between parameter configurations and task performance, the platform enables students to develop intuitive and systematic insights into RL algorithms and decision-making mechanisms. In addition to its instructional value, the scheme demonstrates strong technical feasibility for addressing real-world robotic manipulation challenges, providing a scalable foundation for broader applications. This work offers an innovative approach to bridging theoretical RL education and engineering practice, supporting the development of interdisciplinary talent and the construction of advanced virtual simulation laboratories. Future work will extend the platform to additional manipulation scenarios and control strategies, further strengthening its role in robotics education and research.

Agricultural drought has a negative impact on the development of agricultural production and even poses a threat to food security. To reduce disaster losses and ensure stable crop yields, accurately predicting and classifying agricultural drought severity based on the standardized soil moisture index (SSMI) is of significant importance.

An agricultural drought prediction model, GCN-BiGRU-STMHSA was proposed, which integrated a graph convolutional network (GCN), a bidirectional gated recurrent unit (BiGRU), and a multi-head self-attention (MHSA) mechanism, based on remote sensing data. In terms of model design, the proposed method first employed GCN to fully capture the spatial correlations among different meteorological stations. By utilizing GCN, a spatial graph structure based on meteorological stations was constructed, enabling the extraction and modeling of spatial dependencies between stations. Additionally, a spatial multi-head self-attention mechanism (S-MHSA) was introduced to further enhance the model's ability to capture spatial features. For temporal modeling, BiGRU was utilized as the time-series feature extraction module. BiGRU considers both forward and backward dependencies in time-series data, enabling a more comprehensive understanding of the temporal dynamics of agricultural drought. Meanwhile, a temporal multi-head self-attention mechanism (T-MHSA) was incorporated to enhance the model's capability to learn long-term temporal dependencies and improve prediction stability across different time scales. Finally, the model employed a fully connected layer to perform regression prediction of the SSMI. Based on the classification criteria for agricultural drought severity levels, the predicted SSMI values were mapped to the corresponding drought severity categories, achieving precise agricultural drought classification. To validate the effectiveness of the proposed model, the global land data assimilation system (GLDAS_2.1) dataset and conducted modeling and experiments was utilized on five representative meteorological stations in the North China Plain (Xinyang, Gushi, Fuyang, Huoqiu, and Dingyuan). Additionally, the proposed model was compared with multiple deep learning models, including GRU, LSTM, and Transformer, to comprehensively evaluate its performance in agricultural drought prediction tasks. The experimental design covered different forecasting horizons to analyze the model's generalization capability in both short-term and long-term predictions, thereby providing a more reliable early warning system for agricultural drought.

Experimental results demonstrated that the proposed GCN-BiGRU-STMHSA model outperforms baseline models in both SSMI prediction and agricultural drought classification tasks. Specifically, across the five study stations, the model achieved significantly lower mean absolute error (MAE) and root mean squared error (RMSE), while attaining higher coefficient of determination (R²), classification accuracy (ACC), and F₁-Score (F₁). Notably, at the Gushi station, the model exhibited the best performance in predicting SSMI 10 days ahead, achieving an MAE of 0.053, a RMSE of 0.071, a R² of 0.880, an ACC of 0.925, and a F₁ of 0.924. Additionally, the model's generalization capability was investigated under different forecasting horizons (7, 14, 21, and 28 days). Results indicated that the model achieved the highest accuracy in short-term predictions (7 days). Although errors increase slightly as the prediction horizon extends, the model maintained high classification accuracy even for long-term predictions (up to 28 days). This highlighted the model's robustness and effectiveness in agricultural drought prediction over varying time scales.

The proposed model achieves superior accuracy and generalization capability in agricultural drought prediction and classification. By effectively integrating spatial graph modeling, temporal sequence feature extraction, and self-attention mechanisms, the model outperforms conventional deep learning approaches in both short-term and long-term forecasting tasks. Its strong performance provides accurate drought early warnings, assisting agricultural management authorities in formulating efficient water resource management strategies and optimizing irrigation plans. This contributes to safeguarding agricultural production and mitigating the potential adverse effects of agricultural drought.

In the context of new engineering construction, cultivating students’ ability to solve complex engineering problems has become a focal point of engineering education reform. To help students master robotics and computer vision, a multidisciplinary and innovative experimental teaching program has been designed and developed for a tomato-like intelligent harvesting robot, specifically targeting the field of agricultural robots.

The methodology adopted in this experimental teaching program follows the paradigm of "problem discovery, scheme design, problem-solving." Initially, detailed challenges faced by intelligent harvesting of tomato-like fruits were scrutinized. Ensuring freshness and flavor requires staged harvesting and quick marketing. The long production cycle and short harvesting window of cherry tomatoes necessitate manual intervention. In addition, diverse growth postures of cherry tomato clusters and variations in fruit ripeness pose challenges for harvesting robots performing "fruit picking" tasks. These challenges lead to low harvesting efficiency and difficulties in difficulties in effective graded harvesting. To address these issues, a robot vision detection scheme based on cascade vision is proposed. This scheme achieves precise identification, ripeness detection, and positioning of tomato-like fruits. During the prediction stage of the YOLOv5 network, an additional detection branch is introduced to handle object detection and maturity grading of tomatoes simultaneously. By excluding non-target objects during data annotation and filtering for non-target objects, the computational burden on the network is reduced, thereby improving harvesting efficiency. Additionally, MobileNetv3 is introduced to classify the positional relationship between the fruit and its stem, guiding the end effector to approach the target fruit at the correct angle. To accurately and reliably locate the three-dimensional position of cherries, a cascaded vision-based localization detection method is proposed, combining the advantages of individual single-modal methods and three-dimensional point clouds. Ultimately, a meticulously crafted cherry tomato dataset is constructed, and hand-eye calibration of the robot platform is performed. Three robot harvesting experiments were conducted using cherry tomatoes to validate the effectiveness of the proposed scheme.

The tomato detection and maturity grading tests demonstrate high average recognition accuracy in detecting the ripeness of single tomatoes and clustered targets at various ripeness stages. Incorporating MobileNetv3 led to the clear classification of the positional relationship between fruit and stem, effectively guiding the end effector toward precision. In the final robotic grasping experiment, compared to fixed-angle harvesting, a notable enhancement in harvesting efficiency is observed, along with a reduction in the average time taken per fruit harvested. Consequently, the designed harvesting scheme for tomato-like fruits, based on cascaded visual detection, provides crucial technical support for achieving graded and staged harvesting, significantly improving overall harvesting efficiency.

The design and experimental teaching process of the visual solution enhances students’ understanding of theoretical knowledge in machine vision and deep learning while improving their practical skills in operating robot platforms can be enhanced. This helps cultivate students’ ability to independently solve engineering problems and innovate, enhances their interest in artificial intelligence, and lays a solid foundation for nurturing automation-related talents with interdisciplinary competence.