The track system is inherently cumbersome. Let’s think out of the box.
A: Make the unit independent. It is an autonomous robot.
B: Instead of tracks to give it position info, it uses GPS. Standard GPS is only good to about 3 meters on a good day, but with 8 channel receivers and using raw data, you can have a base station that is fixed, but transmits it’s current location, as calculated by the current constellation of satellites. The robot receives this info and uses it to correct what the same constellation is telling it. This gives about 10 cm accuracy. If you need more info, you can pull phase information out of the carrier signal to get sub cm. accuracy. However for this to work, losing the signal even briefly requires you to go to a known location to re-acquire position.
Alternate position system:
A theodolite measures angles with accuracy as good as 0.1 minutes or 6 arc seconds. (vernier measure) If this can be done electronically, then you could place several corner reflectors around a field, and have a a rotating laser measure the angles between reflectors. 1 degree at 100 yards is about 5 foot error. 1 minute is 1/60 of this or about a an inch. This degrades as the angular difference gets smaller, but with a reflector on 4 corners, you have a lot of redundancy. This in principal allows a hectare to be managed by one unit.
For most uses this is adequate. For watering you have a problem, in that this system does not handle hoses gracefully. I don’t think coming back to fill up with water every few hundred gallons will work. However, if you zone it appropriately, you may be able to make a conventional sprinkler system work for this. Rather than bury the lines, hang them from an overhead trellis. Google “hop yard” to see what I mean.
The present system isn’t good at identifying weeds. This is especially problematic as plants expand into their surroundings. (Consider a squash with tendrils spreading over 10 feet.
System needs a camera, along with software such as “snap leaf” to identify the plant.