I suspect this is a very difficult problem however just because of how varied individual plants can be... telling plants apart can be hard even for trained humans.
Actually, it's not. Despite all the hype surrounding deep learning, tasks like telling similar looking plants apart is precisely what's it's good at. In fact, it will probably be better than most humans at this task. Moreover, it's simple to implement. You take off the shelf model (e.g. ResNet-50) trained on ImageNet, and finetune it on a dataset of specific plants you want it to recognize (~1000 examples of each plant). Then you can run that model on an iPhone (or even on much cheaper hardware)
A much harder problem is to make the robot move through the terrain without getting stuck. But even there, looking at what Boston Dynamics is doing, it seems like we are not that far off.