New AI Helps Robots Spot Tiny Apple Parts in Orchards

Growing apples is harder than it looks. Each spring, apple trees produce far more tiny fruits than they can support. To get big, juicy apples, growers have to remove some young fruits—a process called thinning. Today, that's often done by hand, which is slow, expensive, and depends on a shrinking pool of seasonal farmworkers. This new research aims to change that.

Scientists trained a vision-language AI model to recognize three tiny parts of baby green apples: the calyx (the flower remnant on top), the fruitlet body, and the peduncle (the stem). They used 600 high-resolution photos from commercial apple orchards and tested the model on orchard robots. It scored 95% accuracy on the calyx, 98% on the fruitlet, and 85% on the stem—with an overall score of 93%. That level of precision matters because robots need to know exactly where to cut or spray.

The clever part is size and speed. The full model is only about 130 megabytes—roughly the size of a few smartphone photos—and it can analyze an image patch in milliseconds. It runs on small, energy-efficient computers like those already used in drones and farm robots. That means the "brains" can sit right on the machine, without needing a constant internet connection to a powerful data center.

There are still hurdles. The model was tested on just two apple varieties and a small dataset, so real orchards with different lighting, weather, and tree shapes may trip it up. But the researchers made the code public, so others can build on it. If robotic thinning becomes reliable, farms could save huge amounts of labor and money—and possibly keep apple prices from climbing.