We started a discussion in a Signal group, but wanted to move it here for more involvement. The proposal on the table is to use own source hardware and software to hack together an autonomous weeding machine, maybe using AI to recognize weeds versus crop plants.
We had a few existing examples posted here before.
I was the one proposing it on Signal. I’m personally farming in eastern Europe and interested.
Also, we’re using last century equipment, sometimes middle of last century…
Some relevant projects I found:
[1]: GitHub - AutoRoboCulture/Nindamani-the-weed-removal-robot: Nindamani …
[2]: GitHub - BrandonAllan/Weeding_robot: The Weeding Robot Project is a …
[3]: A Development Platform for Autonomous Weeding. - GitHub
[4]: anida16/Autonomous-InterRow-Weed-Removing-Robot - GitHub
[5]: GitHub - rahularepaka/Project-Cyclops: Weed Detection and Laser …
[6]: Agroecology-Lab/Open-Weeding-Delta: Open Weeding Delta - GitHub
[7]: GitHub - brycejohnston/awesome-agriculture: Open source technology for …
At first, I read “wedding robots” and my mind went in weird directions. :)
If small crawling robots could replace all functions of tractors (incl herbicides/pesticides), then we could really see an efficient agricultural revolution!
I think one of the keys is that they have to be able to be constructed relatively cheaply and with off-the-shelf components. If it costs $5000 and you need an engineering degree to build the final product it isn’t really a leveler, more of a toy for wealthy land owners.
I once considered building one, but never got started because my agriculture took a turn away from field cultures towards trees / bushes (cherry and sea-buckthorn). Now, sadly, I’m involved with far more destructive robotics projects than weed-whacking. :( If the RF stops trying to conquer places and drones aren’t needed to oppose that insanity, maybe I’ll get back to peaceful projects.
Some notes:
-
imaging: I strongly recommend Raspberry Pi and libcamera, can provide directions and sample code (C++) if anyone needs
-
navigation: I considered a tall mast (fishing rod / plastic pipe / other) made rigid with shroud lines and a downlooking camera, information plates with QR or Data Matrix codes embedded in ground (find the information plate by contrast, pass to decoding algo, get reference point coordinates out)
-
weed detection: hardest part, I never got there
-
weed destruction: sensible options seemed like 1000 W of LED light focused by a Fresnel lens, and pressure spraying with water (at the cost of throwing up mud at the camera)
-
motors: I considered extremely slow geared stepper motors with bicycle sprockets and chains, ordinary 4 amp industrial stepper drivers
-
wheels: I considered bicycle rear wheels without a freewheeling mechanism*
-
steering: I considered bicycle front forks, stepper motors and toothed belt (oversized 3D printer parts)
(* actually, a freewheeling mechanism might be handy to retrieve a broken robot, but then you also need brakes)
Basically, it would have been 2 bikes in parallel, rear wheels driven by steppers, front wheels steered by steppers. No brakes. Tall mast, navigation camera at mast top, instrument camera on a movable arm.
-
Using a hyperspectral camera (basically like IR spectroscopy, but done to each pixel in an image) to identify weeds and pests via their chemical makeup would be more expensive but probably more reliable than trying to get AI to identify them based on their visual appearance.
