Abstract
Effective and efficient agricultural manipulation and harvesting depend onaccurately understanding the current state of the grasp. The agriculturalenvironment presents unique challenges due to its complexity, clutter, andocclusion. Additionally, fruit is physically attached to the plant, requiringprecise separation during harvesting. Selecting appropriate sensors andmodeling techniques is critical for obtaining reliable feedback and correctlyidentifying grasp states. This work investigates a set of key sensors, namelyinertial measurement units (IMUs), infrared (IR) reflectance, tension, tactilesensors, and RGB cameras, integrated into a compliant gripper to classify graspstates. We evaluate the individual contribution of each sensor and compare theperformance of two widely used classification models: Random Forest and LongShort-Term Memory (LSTM) networks. Our results demonstrate that a Random Forestclassifier, trained in a controlled lab environment and tested on real cherrytomato plants, achieved 100% accuracy in identifying slip, grasp failure, andsuccessful picks, marking a substantial improvement over baseline performance.Furthermore, we identify a minimal viable sensor combination, namely IMU andtension sensors that effectively classifies grasp states. This classifierenables the planning of corrective actions based on real-time feedback, therebyenhancing the efficiency and reliability of fruit harvesting operations.