Recipe: autonomous row-crop weeder

Problem statement

A 5-acre row-crop field (vegetables, beans, brassicas) has weed pressure that demands daily mechanical cultivation between hand-weeding sessions. The labor cost of daily weeding is the limiting factor for chemical-free farming at this scale. An autonomous machine that drives between rows, identifies weeds, and disrupts each one — repeatable, daily, low-supervision — is the right intervention.

Architecture

[Acorn rover platform]
  └─ 4× hub motors, independent steering (the wheels)
  └─ Pixhawk-class autopilot or low-level MCU (the wheel-control layer)
  └─ NVIDIA Jetson Orin Nano (the perception+planning brain)
       ├─ Intel RealSense D435i (downward-facing canopy view)
       ├─ Forward-facing RPLiDAR (obstacle avoidance)
       ├─ u-blox ZED-F9P GPS-RTK + base station (cm-precision row-following)
       └─ ROS2 graph:
              perception_node:    YOLOv8 inference → weed bounding boxes
              localization_node:  GPS+IMU+wheel-odometry EKF → robot pose
              planner_node:       row-following with obstacle avoidance
              actuator_node:      target-coords → stepper arm motion → tool fire
  └─ Stepper-driven X-Y arm with end-effector (hoe blade / flame / laser)
  └─ Solar panels (300–600 W) on roof
  └─ LiFePO₄ pack with smart BMS

The vision pipeline (the hardest part)

1. Acquire RealSense color frame at canopy distance (~30 cm above soil).
2. Preprocess in OpenCV: white-balance, perspective-correct to a top-down view of the row.
3. Inference with YOLOv8 trained on a custom weed-vs-crop dataset (5,000+ labeled images of this farm’s crops at multiple growth stages).
4. Per detection: map pixel coordinates back to ground-plane coordinates using depth and known camera extrinsics → physical (x, y) of the weed in the rover’s local frame.
5. Hand off target list to the actuator node, which converts (x, y) into [[stepper-motor|stepper]] motion and fires the tool when arm is positioned over target.

Anti-disaster constraints

• No-go on uncertain detection — confidence below threshold = skip the weed, log it for human review. False positives killing crops is unrecoverable.
• Tilt-stop at >15° pitch or roll (IMU watchdog).
• Emergency stop on any LiDAR detection within 2m at the front cone.
• Max-speed cap of 1 m/s — fast enough for productivity, slow enough that humans in the field have time to react.
• Visible operator presence required during early deployment — supervised autonomy until the model has been validated on this farm’s weed flora.

What this recipe does NOT solve (the hard remaining work)

• Custom dataset collection and labeling — weeks of labor per farm; the actual bottleneck.
• End-effector mechanical engineering — the cultivation tool itself (hoe, flame, laser) is its own design problem.
• Headland turning — turning the rover at row-end requires either GPS waypoints or vision-based turn detection; non-trivial.
• Regulatory / insurance reality — autonomous machinery in unfenced fields is not a settled policy area in most jurisdictions.

Cost

• [[acorn-rover|Acorn rover]] kit + chassis: $5,000–10,000
• [[nvidia-jetson|Jetson Orin Nano]] + RealSense + LiDAR + RTK-GPS: ~$1,000
• [[stepper-motor|Stepper]] arm + end-effector: $300–800 depending on tool
• Custom dataset labeling labor: 100–200 hours
• Total: ~$7,000–15,000 in parts, plus labeling labor.

See also

Auto-generated from this entry’s typed relations: frontmatter, grouped by relation type so the editorial signal isn’t flattened.

• Combines: [[acorn-rover]] · [[nvidia-jetson]] · [[rgbd-camera]] · [[yolo]] · [[opencv]] · [[ros2]] · [[gps-rtk]] · [[lidar-rangefinder]] · [[imu-mpu6050]] · [[stepper-motor]] · [[solar-charge-controller]] · [[lithium-bms]]