For Tree-Pruning Robots, Older AI 'Eyes' Beat Trendy New Ones

A robot that prunes trees needs to answer two questions about every tiny dot in its camera view: is this part of a tree, and how far away is it? Researchers at a New Zealand university built a controlled experiment to find out which AI image-reader — called an 'encoder' — does that job best. They kept everything else identical and swapped only the encoder, testing several families of designs.

The results went against the hype. Convolutional and hybrid encoders — older, simpler designs — came out on top. Plain 'transformer' models, the same family of design that powers chatbots like ChatGPT, mostly fell apart when trained from scratch on this task. And model size turned out to be a bad predictor of quality: a small model with roughly 25 million settings outranked models a hundred times larger. More settings, it turns out, does not mean more skill.

There was more good news. Segmentation (telling tree from not-tree) and depth (judging distance) ranked the models almost identically, meaning the two jobs don't fight each other — one encoder can handle both. That's efficient. But the team also caught a hidden trap: several models 'cheated' by labeling everything as a tree, which looks fine under one common scoring method but fails under a stricter test that checks edges. That matters, because a robot that thinks the whole world is a tree could crash into things.

For anyone outside the lab, the takeaway is this: the smartest choice isn't always the flashiest. Careful measurement beat reputation, and the winning designs were also the cheaper ones to run. That's good news for the future price of agricultural robots, drones that map forests, and any camera that needs to understand both what it sees and how far away it is.