Autonomous Agro-Tech Rover

A 4WD differential-drive rover designed to operate without cloud connectivity in the field. RTK-GPS waypoints define the mission, a 2D LiDAR provides obstacle avoidance and row-following, and an onboard multispectral pipeline computes NDVI and runs an EfficientNet-Lite classifier on a Coral TPU for pest and disease detection. Telemetry is buffered locally and synced via MQTT when connectivity returns.

Smallholder and mid-scale farms can't justify the cost of commercial precision-agriculture platforms, and most existing rovers depend on continuous cloud connectivity that simply doesn't exist in the field. The system needed to operate fully offline, survive dust and vibration, and still produce per-plant insights worth acting on.

Navigation runs on ROS 2 with an RTK-GPS waypoint follower and a 2D-LiDAR row-detection node feeding a local costmap. Perception is decoupled: a multispectral capture pipeline computes NDVI bands and feeds an EfficientNet-Lite classifier quantized to INT8 on a Coral TPU, keeping inference under 30 ms per frame. Low-level motor control runs on an STM32 over CAN, with a watchdog that fails safely on perception or planner stalls. Telemetry is written to a local InfluxDB buffer and synced via MQTT when uplink is available.

Reliable autonomous traversal of row crops at ~2 acres/hr with 99.5% obstacle-avoidance success in field trials. The on-device classifier reaches 91.2% top-1 on a 6-class pest/disease set with sub-30 ms inference. The system runs an entire mission untethered and reconciles cleanly with the cloud dashboard once connectivity returns.