a minimal-hardware approach to weeding
The idea presented here applies only to weed seedlings. Weeds growing from tubers or invasive roots will need to be handled more aggressively, but seedlings, being poorly rooted, are vulnerable to methods that destroy their single meristem. Moreover, after a few years of careful weeding, they are the only type of weed that would persist, except for those growing from runners invading from adjacent land, around the perimeter of the plot, so this method would become gradually more sufficient.
In a nutshell, the idea is to use video imagery to locate seedlings, an expert system (the hard part) to distinguish between desirable seedlings and weeds, and a pulse laser to first make sure it has a clear path to the weed seedling (nothing in the way), focus on the portion of the seedling containing the meristem and then deliver one or more relatively high-energy pulses to heat it sufficiently to render the meristem inert, so that the cells are no longer capable of growth and division. It isn't actually necessary to kill the meristematic tissue outright, just inactivate it, so the higher energy pulses used to accomplish this should not need to be so powerful that they present any danger of fire.
Of course, if the machine carrying out this task maintains or has access to a very detailed map of the plot, which precisely locates and keeps an image archive of every seedling, the next time it passes nearby it can simply check whether the plant appears to have withered, or whether it has recovered and continued growth, in which case it may be time to call in heavier equipment. In this way it can build experience with just how much energy is required to stop the growth of a weed seedling of a particular type at a particular stage in its development. Weeds that survive the surgical approach of the laser can be dealt with by more conventional mechanical methods.
The video system should at least combine a wide-angle view with a telescopic view (needed to distinguish between weeds and desirable seedlings). Either or both might be binocular (stereo), for 3D capability, and the telescopic view in particular would benefit from the use of a sensor that could deliver partial frames very rapidly, to help assess the effectiveness of the laser pulses (how much does the meristem swell within the first tenth of a second?).
I call this a minimal-hardware approach because it involves little more than a pair of cameras, one wide-angle and the other telescopic (two pair for stereo video at both focal lengths) and a laser, on a mount with two degrees of freedom, both rotational, and some means of moving that mount around a plot or field. The real complexity would be in the software that deciphered the video input, deciding which seedlings to zap and which to let live. A high-pressure water jet could be substituted for a laser, but such an arrangement would be more challenging mechanically, because the nozzle would need to either come within a few inches of the seedling or use a significant amount of water to be effective. Too much water applied at high pressure might create other problems, for example encouraging the growth of fungi.
The knowledge necessary to distinguish between seedlings of various species would be an appropriate addition to the RoboEarth project.