Carbon Robotics swaps dozens of crop-specific AI models for one that learns a new field in minutes

A single 'large plant model' trained on 150 million labeled plant images lets farmers define crops and weeds on an iPad, then point a laser weeder at exactly what they want gone.

AI2Day Newsdesk4 min read
Warehouse automation
Share

Key points

  • Carbon Robotics replaced multiple crop-specific AI vision systems with one large plant model trained on 150 million labeled plant images.
  • Farmers can configure the system for a new crop or region in minutes using an iPad app, with no software downloads required.
  • The new model works by comparing plants it sees to examples a farmer tags, rather than assigning fixed "weed" or "crop" labels.
  • Carbon Robotics also released Carbon ATK, a kit that retrofits John Deere 6R, 8R, 8RX and 8RT tractors (2019 and newer) for fully driverless operation.
  • Data-annotation company iMerit labeled over one million plant images to make the foundation model possible.

For years, AI-powered farming machines had a problem: teach one to recognise weeds in a carrot field in Arizona, and it would need a fresh round of training before it could do the same job in a lettuce field one state over. Carbon Robotics, which makes autonomous laser weeders, just changed that.

The company has replaced its collection of separate, crop-specific AI vision models with a single system it calls a large plant model. Think of it like the difference between hiring a specialist for every task and hiring one expert who learns your specific needs on the spot.

The model was pre-trained on 150 million labeled images of young plants collected from farms around the world. That broad base means it already understands how plants are shaped and structured across hundreds of species.

How does a farmer actually set it up?

Setup takes minutes, not days. A farmer opens an iPad app, reviews small thumbnail photos pulled from their own field, and taps a handful to mark as "crop" or "weed."

Those tagged examples go straight into the model, which immediately adjusts what it will target with the laser. No new software. No waiting for a model to retrain overnight. The weeder can zap weeds in an Arizona carrot field in the morning and switch to herbs on a different farm the same afternoon.

Carbon Robotics CTO Alex Sergeev, speaking to The Robot Report, described the shift in how the model reasons. "The main thing that model needs to output," he said, "instead of saying 'Here's my confidence that this is a weed,' it gives you a way to do comparison between this plant and all the plants that the farmer told you are crops and weeds. That's the difference, and that needs to happen really fast."

When the laser removes a weed, the plant material breaks down and returns nutrients to the soil, a small bonus for whatever crop is growing nearby.

What is Carbon ATK?

Carbon ATK removes the need for a driver altogether. It is a hardware kit that bolts onto existing John Deere 6R, 8R, 8RX and 8RT tractors (2019 models and newer) without permanent modifications.

The kit adds cameras and sensors for obstacle avoidance, route planning and position tracking. A tractor fitted with Carbon ATK and one of the company's smart implements, like the LaserWeeder, can run complete weeding missions with no one in the cab. The same tractor can also handle spraying, tillage and harvesting.

Feature Detail
Training data 150 million labeled plant images
Field setup time Minutes (iPad app, small number of tagged examples)
Data-labeling partner iMerit (1 million+ images labeled)
Carbon ATK compatible tractors John Deere 6R, 8R, 8RX, 8RT (2019+)
Software update needed to switch crops None

Common questions

Does this replace a farm worker?

The autonomous tractor handles repetitive field passes on its own, but someone still needs to set up the iPad app and supervise operations. Carbon positions the technology as handling the most time-consuming physical work.

Will this work on any crop?

The large plant model was trained across a wide range of species and geographies, but a farmer still tags a few examples from their specific field before each new deployment, so the system matches local conditions rather than guessing.

© 2026 AI2Day