The problem of automated plant identification has been examined [0, 1] with good results. In [0] the classification is performed on footage from an uncontrolled agricultural setting, and the authors achieved a correct classification rate of 65% on tomato plants. Given that modern industrial agriculture is monoculture you only need to positively detect one variety of plant in an area. To avoid crop destruction due to incorrect categorization the robot could be used as a first pass and restricted to punching plants that are classified with a high degree of confidence. This could still reduce the amount of human labor necessary to tend to the field, but it might not reduce it by a degree that makes the robot a sensible economic solution. It may fail to reduce the labor needed, but I do not think that is likely.

I wonder how the performance of plant recognition degrades over the course of the crop cycle, if at all. We see in the video that some leaf matter remains after the weed is driven into the ground, and this leftover matter will influence classification during the robot's next pass. Unless the leaf matter is removed between passes it will accumulate throughout the crop cycle.

In [0] the researchers note that one reason they chose to deal with seedlings is that there is relatively little plant leaf occlusion at that stage. In [1] an end user is relied on to take a photo with a light, untextured, background and does not at all deal with partial plant leaf occlusion. That paper is not applicable to the uncontrolled field scenario.

I also found [2] while writing this comment, but have yet to read it. It boasts even better results (>90% correct for corn, 73.1% correct for tomatoes in an uncontrolled setting)than [0]

[0] http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.42....

[1] http://neerajkumar.org/papers/nk_eccv2012_leafsnap.pdf

[2] http://www.mdpi.com/1424-8220/11/6/6270/pdf