If it works - great (spectral analysis is pretty cheap, and fairly light in terms of the computing power needed). But as your article states, even among the real experts that method is debatable. I imagine (having no knowledge of their field) that their issues are the same as the ones I brought up - variation in the same species, limited variability among species, and the fact that many weed species are the wild variants of the crop.
Quite possibly. There's only so much useful speculation to be done here regarding detection. But apart from the detection issue, there are other problems. These little guys would need to be capable of rapidly traversing soft and uneven terrain, and of killing weeds by some as yet undetermined method. They would also need a sufficient power source. Some people, sometime - my guess is quite a few years from now - may do the R & D and iron out the wrinkles using a series of prototypes. For the time being, this critter looks very bulky and prohibitively expensive; but I can see it coming. It's an idea too good to ignore.