Written by 
Battling weeds can be expensive, labor intensive and use large amounts of chemicals. To help make this easier [NathanBuilds] has developed V2 of his open-source drone weed spraying system, complete with automated battery swaps, herbicide refills, and an AI vision system for weed identification.
The drone has a 3D printed frame, doubling as a chemical reservoir. V1 used a off-the-shelf frame, with separate tank. Surprisingly, it doesn’t look like [Nathan] had issues with leaks between the layer lines. For autonomous missions, it uses ArduPilot running on a PixHawk, coupled with RTK GPS for cm-level accuracy and a LiDAR altimeter. [Nathan] demonstrated the system in a field where he is trying to eradicate invasive blackberry bushes while minimizing the effect on the native prairie grass. He uses a custom image classification model running on a Raspberry Pi Zero, which only switches on the sprayers when it sees blackberry bushes in the frame. The Raspberry Pi Global Shutter camera is used to get blur-free images.
At just 305×305 mm (1×1 ft), the drone has limited herbicide capacity, and we expect the flights to be fairly short. For the automated pit stops, the drone lands on a 6×8 ft pad, where a motorized capture system pulls the drone into the reload bay. Here a linear actuator pushes a new battery into the side of the drone while pushing the spend battery one out the other side. The battery unit is a normal LiPo battery in 3D-printed frame. The terminal are connected to copper wire and tape contacts on the outside the battery unit, which connect to matching contacts in the drone and charging receptacles. This means the battery can easily short if it touches a metal surface, but a minor redesign could solve this quickly. There are revolving receptacles on either side of the reload bay, which immediately start charging the battery when ejected from the drone.
Developing a fully integrated system like this is no small task, and it shows a lot of potential. It might look a little rough around the edges, but [Nathan] has released all the design files and detailed video tutorials for all the subsystems, so it’s ready for refinement.