Generating realistic pruning solutions for automated grape vine pruning using graph neural networks

In our prior work we showed that graph neural networks (GNNs) can be trained to generate pruning solutions that could direct robotic pruning robots to perform automated cane pruning of wine grape vines. That study introduced the feasibility of the technology but also showed that there were many open questions and issues with the research results that needed to be addressed. In this study we address some of these questions. For example, we answer the question of how would a model like this perform on real vine architectures compared with pruning solutions from real experienced pruners. Our most notable contributions include moving away from a per-cane classification model that attempts to define a single perfect pruning solution, to a model that ranks multiple good solutions and picks the best one. We addressed a key limitation of the previous training data by moving away from synthetic vine architectures to realistic ones recorded from real vines and using pruning solutions collected by expert pruners as our ground-truth. Our primary goal was to show that learning by example using a GNN-based model was a viable approach to automated pruning, even when compared with experienced pruners. We showed robust performance from our model by training on a dataset of 90 pruning solutions generated by expert pruners in the 2022 season, and testing our performance on 117 pruning solutions from an independent set of pruners from the 2021 season. The model was able to correctly score all the pruning solutions from the 2021 dataset as good to very good and none of the expert solutions were classified as poor.