Development of a decision-making algorithm to identify picking strategies for robotic mushroom harvesting

Button mushrooms (Agaricus bisporus) are harvested manually around the world to preserve high quality for the fresh market due to their sensitive skin. The United States mushroom industry, as one of the massive global mushroom producers, has confronted high production costs and labor shortages due to labor-intensive mushroom harvesting. The need for alternative solutions, such as robotic harvesting that imitates common manual picking strategies, specifically bending, twisting, and lifting, is becoming more and more important for the industry. This research seeks the development of a decision-making strategy to pick mushrooms with an effective picking sequence and bending orientation. The YOLOv5s model was utilized to detect and locate mushrooms. The maturity of the mushrooms and their clusters was identified using a pixel-based threshold. Bending distance was determined through geometric analysis, and spatial clustering was performed with the DBSCAN method. The spatial arrangements of the mushrooms were uncovered, a picking sequence was assigned, their bending directions were determined, and a human picking test was conducted to validate the approach. The results demonstrated a mean average precision (mAP) of 97% for the YOLOv5s model, R² of 97% for maturity detection (pixel count vs. distance), bending distance defined by a product of a bending coefficient (k) of 0.21 ± 0.09 and the cap diameter of the targeted mushroom, and a weighted silhouette coefficient of 0.28 ± 0.13 for clustering performance. The developed picking strategy successfully identified spatial arrangements of the targeted mushrooms under various clustering conditions. All mushrooms detected during the model evaluation were successfully provided with the correct picking sequence and bending orientation (100%). These findings suggest that the proposed picking strategy can be employed to automate robotic harvesting systems in the future.