I've been thinking about this, the application of robotics to horticulture on a scale large enough to replace (some significant portion of) conventional agriculture, for a very long time, and I'm prone to glossing over points that may not seem at all obvious to others.
For example, if these robots that I've been talking about aren't engaged in tillage, what are they doing? That remains an open question, since there are undoubtedly useful techniques I haven't yet thought of, but, for an idea of what might be possible, consider what gardeners can accomplish with their own bare hands and short-handled tools. That's the scale of manipulation I have in mind, working with individual plants and the spaces into which they're to be inserted.
Would such robots have human-like hands? Only in the vaguest sense; they're likely to have manipulators with opposable, finger-like appendages. Would they stir the soil like a gardener does with a trowel? Maybe. Would they use something like snips to do pruning? Probably, although there might be a better approach to pruning than mechanical snips, like a high velocity water jet (such as are used to cut steel in some industrial settings).
It isn't necessary, nor even desirable, to exactly replicate the set of techniques used by a gardener. Such machines would need a repertoire of techniques sufficient to manage a garden, but while some of their techniques might seem quite familiar, others might be quite beyond the capability of a human gardener.
For example, if a machine were able to identify a weed seedling early enough, it need only destroy the seedling's meristem to interrupt the growth of a weed. This requires very little energy, and might be accomplished by a precisely targeted, high velocity water droplet. Using this method, a machine might deal with several weed seedlings per second, limited only by the speed with which it could identify them and reoriented the nozzle, all without any disruption to surrounding plants.
More tenacious weeds that sprout from roots could be pulled out, except that they sometimes break off just below the soil surface, and their roots may pass below plants you'd rather not disturb. An option would be steam injection, through a tube inserted next to the stem. Another option would be coring, removing a cylinder of soil around the stem to a depth of a few inches. Yet another option would be to use electrical current to heat the weed. These are all techniques that a gardener might use, but, except for grabbing ahold of the base of the stem and pulling the plant out, they aren't common.
Compared with weeding, seed planting would be relatively simple. On the other hand, transplanting seedlings started elsewhere would be more challenging, although mechanical systems for this purpose probably already exist and could be used as a model.
Dealing with mid-season issues, like insects and nematodes, microbial infections, plant nutrient deficiencies, and so forth, is hugely complicated, and will require considerable development effort. But small-scale machines have an advantage in that they can deal very specifically with the effected leaf, plant, or location, and also in that, because they would revisit each location frequently, they should be able to catch problems early.
Harvest is also somewhat complicated, since each crop type presents its own set of challenges. What works for wheat doesn't work so well for maize. What works for tomatoes won't be sufficient for pumpkins. Hardware attachments may be needed in some cases.
This vision isn't a fantasy, but there's a lot of work to be done.