Vineyards change quickly as the season progresses. One week the clusters are barely visible, and a few weeks later the canopy fills with fruit. Keeping track of how much is growing—and where—can be a challenge for growers and an ongoing task for technicians and researchers. In this study, Giuseppe Giorgio Rinonapoli, Rosa Pia Devanna, Giulio Reina, Fernando Auat Cheein, and Annalisa Milella explore a way for a robot to take on this responsibility by estimating grape yield directly from RGB-D images collected in the field.
The authors combine two complementary ideas: a deep learning model that highlights grape regions in each image, and a depth-based process that separates individual bunches even when they overlap or blend into the foliage. Together, these steps help the robot form a clearer picture of how many clusters are present and where they are located. The examples shown in the paper—particularly on pages 8 to 10—illustrate how this method copes with shadows, uneven light, and grape bunches sitting at different depths in the canopy.
What makes the approach practical is that it was tested not in ideal laboratory conditions, but in a real commercial vineyard, using thousands of images captured by the Polibot platform as it moved along the rows. The system consistently identified clusters across a variety of situations, producing estimates that closely followed manual counts.
For farmers, this type of tool could offer a more dependable way to understand crop development without needing to survey large areas by hand. And for the AgRibot community, it demonstrates how pairing modern segmentation networks with depth information can turn everyday field data into meaningful yield insights—supporting better planning throughout the growing season.
If you’re curious to see the full workflow and the visual results, the complete paper is available on Zenodo.