A research team led by Prof. HUANG He at Hefei Institutes of Physical Science, Chinese Academy of Sciences, has developed a vision-based system to help weeding robots identify and locate weeds more accurately in fields of Chinese medicinal herbs.
The research was published in Artificial Intelligence in Agriculture.
Weeding in Chinese medicinal herb fields is still largely done by hand, making it labor-intensive and costly. The task is particularly challenging in fields of Paeonia lactiflora Pall., a perennial herb that grows in dense clumps. Crops and weeds often look similar, plants overlap, and changing light conditions make it difficult for robots to distinguish weeds from crops using conventional cameras.
To improve weed recognition, the team developed a robotic weeding platform and built an image dataset containing about 47,000 annotated plant samples. Using both color images and depth information, the researchers developed a detection framework called FW-DETR. Unlike conventional vision systems that rely mainly on color images, the framework also uses depth information to distinguish plants based on their spatial structure.
The system combines information from different image scales to capture both the overall layout of plants and their finer details. This helps it distinguish weeds from crops even when plants overlap or are surrounded by complex backgrounds. The team also developed a method to determine where the robot should target each weed, taking into account its shape and the extent to which it is obscured by other plants.
Tests under field conditions showed that FW-DETR achieved an average precision of 84.4% at an intersection-over-union threshold of 0.5, outperforming several existing detection models, including RT-DETR and YOLOv12-L. The system could also locate weeding targets with millimeter-level accuracy, providing the precision needed for robotic weeding.
The findings offer a practical basis for developing more accurate weeding robots for Chinese medicinal herb cultivation. The approach could also be adapted for weed management in other medicinal herb fields and agricultural crops, according to the research.

Technical workflow for weed detection in Chinese medicinal herb fields. (Image by SUN Youqiang)