Manual fruit picking is labor-intensive and can damage fruit. Fully mechanized picking is efficient, but it also risks fruit damage. Therefore, semi-automated tools are needed to improve bitter orange picking. This paper presents a smart manual picker designed to facilitate picking while predicting fruit maturity based on picking force as well as various chemical and physical parameters using machine learning (ML). The study methodology consists of five stages: (1) manufacturing the smart picker, (2) picking 50 bitter orange samples, (3) measuring the characteristics of the bitter oranges in the laboratory, (4) training different ML models, and (5) identifying the most accurate model for predicting fruit maturity. The results indicate that as fruits mature, their weight, CIE-L*a*b* values, and pH levels increase, while picking force and hardness decrease. Notably, picking force exhibited a strong correlation (93.5%) with maturity compared to other physical parameters. The Kruskal–Wallis test also showed that the relationship between picking force and bitter orange physical parameters, including weight, CIE-L*a*b*, pH, and hardness, was statistically significant. The extreme gradient boosting (XGBoost) model achieved the highest training accuracy (100%), outperforming stacking (99.91%), random forest (91.17%), and gradient boosting machine (89.08%) on all evaluation metrics. However, the stacking model is considered better, even though XGBoost achieved 100% training accuracy, as the former showed a better balance between training, testing, and validation. This study contributes to improving bitter orange quality by accurately predicting maturity through data collected from the smart picker.
