The ripeness model achieved mAP@0.5 of 0.937 with strong true positive rates for ripe and unripe classes. The multi-class foliar pathology model achieved mAP@0.5 of 0.636 across challenging disease categories.

Dockerized deployment on Raspberry Pi 5 sustained an average inference latency of about 42ms, meeting real-time operation demands for a moving robot platform.

LoRa mesh operation maintained packet delivery above 98% in dense greenhouse conditions. Under simulated link failure, SDN-guided failover completed in around 165ms while preserving critical command flow.

Spatial clustering for spray targeting reduced redundant arm movements and helped reduce chemical use by about 30%, improving sustainability and reducing intervention overhead.

The RP confirms that combining perception quality, actuation precision, and SDN-based continuity is more valuable than optimizing any single module in isolation. This is the main reason the platform remains functional during realistic greenhouse disturbances.