#番茄剪枝机器人

开发基于YOLOv8n-BiFPN-WIoU的轻量模型，实现温室番茄自主采摘机器人实时运行。

提出3MSP2方法，实现温室大果番茄从簇级排序到单果姿态估计的序贯采摘规划。

设计设施番茄智能采摘平台，融合改进YOLO v5与HSV实现识别定位与采摘。

北京市智慧农业创新团队柴秀娟研究员团队提出基于VR的专家经验捕获与行为建模框架，融合高保真虚拟环境与人类作业动作知识，显著提升机器人在复杂农业场景中的自主决策能力，相关成果入选 IJCAI 2026 国际人工智能联合会议并将在德国会议期间进行现场展示，也是中国智慧农业团队在国际 AI 顶会上的重要亮相。技术上，机器人先用RGB-D相机高精度扫描真实温室，在Unity引擎中利用视觉感知与3D建模算法重建番茄空间分布及果实成熟度信息；专家知识虚拟化方面，多位专家穿戴27节点高精度动捕设备+VR头盔在高保真虚拟温室中执行多种难度采摘任务；机器人行为智能化方面，用高质量专家数据训练多任务强化学习模型，再通过逆运动学算法将人类动作映射到机械臂运动空间。综合评估显示该"专家数据驱动"方法在收获效率、避障能力、任务成功率、运行稳定性等核心指标显著优于现有主流方法，特别在遮挡严重、果穗密集场景下表现出接近人类专家的适应性与低损伤率（注：目前以番茄采摘为验证场景）。

赵丽成、卢鑫羽、吴茜、任妮、周玲莉、程雅雯、胡安琦、戚超(江苏省农业科学院农业信息研究所/淮阴工学院)提出基于改进YOLOv10的轻量级串番茄成熟度目标检测模型LampCT-YOLO:在主干网络引入SegNeXt注意力机制,基于梯度的全局通道重要性方法LAMP通道剪枝显著压缩模型体积和计算复杂度。LampCT-YOLO对红熟早/中/晚期mAP分别达84.6%/89.5%/88.4%,应用LAMP通道剪枝后参数量和计算量分别减少63.07%和50.06%,推理速度提升23.1%;部署至自主研发的果蔬巡检机器人在边缘设备NVIDIA Jetson AGX Orin上达到9.1帧/s推理速度。

The growth status of tomato plug-tray seedlings varies continuously, with ambiguous boundaries between grades. Existing methods primarily focus on object detection and coarse-grained classification, making it difficult to fully exploit instance-level visual evidence for fine-grained seedling status assessment and interpretable sorting decisions. To address these limitations, this study proposes a detection-aware RGB-D vision–language framework. An RGB-D instance-aware detection model, RDIA-RTDETR, is developed to integrate RGB appearance and depth geometry through bidirectional cross-attention, enhancing the joint representation of leaf morphology, canopy structure, and plug-cell spatial characteristics to improve the detection accuracy of empty cells, weak seedlings, and Strong seedling. A Detection-aware Semantic Bridge (DSB) module is designed to transform object locations, class confidence scores, and detector query features into instance-level conditioning information, establishing a connection between object detection and subsequent semantic reasoning. Furthermore, a Detection-Aware Structured Seedling-status Report Generator (DA-SSRG) is proposed to integrate individual-seedling visual features with detection-bridge features through gated fusion. By incorporating hierarchical discrimination and risk-sensitive constraints, the model performs five-level seedling status grading and sorting decisions while generating corresponding diagnostic reasons and management recommendations. Experimental results show that RDIA-RTDETR achieves an F1-Score of 95.8% and an mAP50 of 97.3%, representing a 4.6-percentage-point improvement in F1-Score over RT-DETR-MM. The complete framework achieves accuracies of 92.73% and 94.17% for fine-grained grading and sorting decisions, respectively, with a Cause Score of 88.2%, a Suggestion Score of 91.5%, and an overall semantic consistency of 93.6%. The missed-identification rates for severely weak seedlings and samples requiring replacement are reduced to 2.3% and 1.0%, respectively. A web-based prototype enables structured report generation for individual plug cells and seedling status summarization at the tray level, providing technical support for abnormality screening, localization of cells requiring seedling replenishment, and fine-grained sorting decisions in plug-tray seedling production.

A closed-loop spatial–temporal–path–execution coordination framework for dual-arm tomato harvesting in constrained shared workspaces。Biosystems Engineering

经济参考报8月7日深度报道江苏省农业科学院打造农业前沿技术集群、加快技术成果从展示品到大田标配转化的最新进展。基因层级研发链条初步成形:湖羊团队靶向抑制肌肉生长关键基因MSTN,大幅提升肉用性能,辐射新疆、甘肃、青海等西部肉羊产区;畜牧所针对部分人群牛奶不耐受问题,聚焦主要致敏原β-乳球蛋白,创制可生产低致敏牛奶的奶牛新种质;种质资源与生物技术研究所赵涵团队利用AI无人机基因测序,一次精准鉴定上百份样本,筛查效率大幅提升,新品种培育周期从传统10年压缩到3年。今年新投用的江苏农业种质资源综合基因库能保存55万份种质资源,满足江苏未来30年农业科技创新需求。智能装备方面,信息所历经6年开发出3代农业机器人,采摘一个番茄从最初的15秒降到9秒,机器人成本从60万元降至20万元以下,2025年获全国遥感高分领域竞赛第一名,采摘机器人获雄安智慧农业大赛二等奖,2026年初农业农村部批准建设的智能采收机器人重点实验室由集萃智能公司联合江苏省农科院等共同打造。冷链流通方面,装备所新技术将果蔬保鲜期从5天延长至15天以上,近5年累计保鲜蔬菜6万多吨,平均减少蔬菜产后损耗12%以上,新增销售额75.4亿元、累计新增利润22.5亿元,李鹏霞团队的冷链流通技术通过美国UC Davis采后研究与推广中心评估,正式上线其全球权威的ProductFactSheets平台。