I am not an expert in robotics (definitely not in pear trees pruning either) but I will try to give some hints to partially answer and also to help reframe the problem a bit. On overall I'll give already an answer: it is most likely possible, but also most likely not convenient.
First things first: in general the rule is that machine learning should be applied when a task can't be automatised otherwise. Before asking if an AI can be trained to solve a task one should always think how to automatise a task using a rule based system. This is important especially because during the process of thinking how to automatise the task you will realise that some steps can be perform without an expert, while others can't be performed without any supervision. Let's brake down your task in subtasks: a system that prune trees should be (at least) capable of:
- Moving between trees
- Selecting branches to cut
- Cut the selected branches
Selecting the branch to cut is probably the step that requires most of the know how and expert supervision and for which a machine learning component might be suitable. Moving instead is a perfect example of a subtasks that could be tackled at different levels. Creating a machine able to anticipate other objects movements and avoid them in real time definitely require to train an AI component, but when you say that the environment is highly structured (trees disposed in a grid) this make me think that maybe some hand coded rules would do the trick without bothering machine learning.
Once you have understood which subtasks your machine should be capable of solving, you can start dig on the theoretical feasibility of them. Following the same order as in the previous paragraph:
Self-driving is a widely studied topics, the algorithms used to train robotic vacuum cleaners to move automatically in a house could be applied straight away to the problem of training an agent to move between trees.
Selecting the right branches involves mostly computer vision. This sub-task should actually be also dived again into sub tasks: detecting branches from other objects and select which ones should be cut. Nevertheless, the field is quite huge and training two models able to perform both actions is, in my opinion, doable.
Cutting a branch is probably harder than driving in this situation, because of the small movements that might be required to reach branches positioned in difficult spots. Anyway, it is possible to train robots to perform fine-grained movements (for a funny example see: Robot learn to flip pancakes).
Again it depends also on how high your expectation are about the final machine/system. Should the system have a surgical precision or could it risks to brake few extra branches in hard situations? Obviously the higher the expectations the harder it would be to make everything work harmoniously.
Last but not least you also need to understand if what you're trying to do is feasible in reality and not just in theory. A big problem when it comes to use machine leaning is that these models can be trained only with huge amount of data.
Train an artificial agent to move in an environment can be done by reproducing the environment in an artificial simulator, which is good news. A single guy with a laptop could potentially do the job.
Collecting data to train a model able to detect branches on photo and then select which one should be cut will be highly tedious, and also expensive because data need also to be labelled by experts. Which means that some experienced people will have to take photos, in the order of tens of thousands at least, and write for each photo (using a software) which are the branches they will cut if they were working for real on that tree. I strongly doubt that a dataset like this already exists.
Training a robotic arm to cut branches will also be challenging. Despite the fact that also in this case simulators could be leveraged, the task is inherently harder and this comes with bigger difficulties (for example in designing a proper reward function if using reinforcement leaning). Concretely this mean more time to spend in research and testing.
Consider also that the success of the final model trained for each subtask would be not guaranteed at all, reason why I said at the beginning that training a system with AI modules would be probably not convenient, and that the best thing is always to try to create a rule based system first.