The edge AI model classifies ripe and unripe tomatoes in real time. When multiple ripe detections appear in one frame, the system selects the highest-confidence instance as the actuation target to ensure deterministic arm behavior.

Instead of directly controlling the arm from fluctuating video boxes, Agri-Sense uses snapshot-based inference. This decouples perception from actuation timing and reduces coordinate jitter during arm movement.

For harvesting, the system computes full (X, Y, Z) coordinates. The Z-axis is estimated via monocular focal-length logic. Then inverse kinematics is solved for the 4-DOF robotic arm to position the end-effector accurately.

The modular 3D-printed gripper executes a torque-limited plucking sequence to avoid damaging fruit and nearby foliage. This directly supports precision harvesting under dense greenhouse conditions.

The same arm switches to spraying mode for pathology treatment. Lesion detections are clustered spatially to produce one optimized centroid spray target, minimizing unnecessary arm motion and reducing chemical usage.

After each harvest or spray action, the system re-captures a validation frame to confirm target completion. If confidence remains low or target status is uncertain, the robot defers repeat action and raises an operator alert instead of applying uncontrolled interventions.