Robotic purification for large-scale seed production of hybrid rice

China is a principal region for hybrid rice cultivation, with an approximate cultivation area of 160,000 ha dedicated to hybrid rice breeding. Ensuring genetic purity, quality, and yield consistency requires effective roguing — the removal of impurity plants — which is a vital step in hybrid rice seed production. This necessitates the rigorous elimination of nonconforming plants. Our work presents a decentralized architecture of multiple unmanned ground vehicles (nUGV) working in concert with one unmanned aerial vehicle (1UAV) to execute this impurity removal task in large paddy fields. Abnormal hybrid rice in hybrid-rice seed fields distinguishes two classes: heterologous and abnormal mature varieties. The digital surface model is used for global localization of impurity plants in the breeding field, the t -SimCLR target detection algorithm is applied for local detection of Abnormal mature varieties, and the motion-coupled mapping algorithm is utilized for local detection of Heterologous varieties. The results indicate that, over the three impurity removal operations (Elongation, Booting, Heading), the IR-robotic system equipped with a cutter outperformed the IR-robotic system equipped with a flamethrower or Laser Feeder in four aspects: impurity removal rate, rice injury rate, true operation rate, and redundancy operation rate. In terms of operational efficiency, under the nUGV-1UAV robot swarms paradigm, the IR-Robotic (UGV) travels at a speed of 3 m/s across an experimental field measuring 239,827.42 m 2 (360.94 mu), with an average Impurity density of 0.0065/m 2 . For the 360.94-mu plot, under normal low-impurity production conditions, the maximum Rice injury rate stands at 0.0881%. When operating in high-impurity traversing mode, the cutter-based system exhibited a higher injury rate of 5.29%. Over the 360.94-mu field, 36 UGVs covered 11,391.97 m (of which 10,554.32 m was effective working distance), consuming 67.15 h in total. For the same 360.94-mu plot, the nUGV-1UAV robot swarms ( n = 36) took 3.48 h to complete. The time taken to complete the same task has been reduced by 94.82%. This study demonstrates a decentralized aerial-ground robotic swarm that addresses key agricultural challenges—including dynamic tasks, unstructured navigation, and environmental uncertainty—through the validated example of automated impurity removal. The system proves to be a feasible, efficient, scalable, and precise solution for ensuring genetic purity in large-scale hybrid rice seed production, thereby providing a clear blueprint for the future of smart and modernized agriculture.