The robot uses an IR-reflective sensor array with PID-based path following to stay centered in greenhouse corridors. A LoRa mesh built on ESP32 nodes captures humidity, temperature, and soil conditions for continuous edge awareness.

Containerized YOLOv11 runs on Raspberry Pi 5 for dual-cycle inference: ripeness classification and foliar pathology detection. The system uses snapshot-based inference to reduce bounding-box jitter and applies a 0.5 confidence threshold for novelty handling.

When multiple ripe fruits are detected, the highest-confidence instance is selected as the actuation target. The system maps detections into (X,Y,Z) coordinates, estimates depth for Z, and uses inverse kinematics to guide the 4-DOF arm and torque-limited gripper for safe harvesting.

For foliar lesions, detections are clustered to compute a single centroid target. The nozzle aligns on (X,Y) while keeping fixed spray distance to avoid contact. This minimizes redundant arm motion and reduces chemical usage.

The robotic arm switches between a modular gripper for fruit plucking and a sprayer nozzle for pathology treatment. This dual-mode actuation converts AI detections into direct physical response without manual intervention.

An SDN control plane (ONOS + OVS) prioritizes intervention packets over routine telemetry. The architecture supports failover from primary links to backup channels using secure VPN-assisted connectivity.

ESP32 nodes use multi-hop LoRa routing so obstructed nodes can relay through neighbors instead of failing silently. The control layer monitors RSSI/SNR trends and can tune communication settings to maintain reliability in humid, high-density foliage environments.

If the model detects unknown symptoms or low-confidence anomalies, a GSM alert is sent to the greenhouse owner with context data, enabling supervised escalation when required.