Carbon Robotics and iMerit Enable Instant In-Field AI Customization

Carbon Robotics is bringing the era of foundation models to agriculture. The company has replaced its collection of crop-specific AI vision models with a single 'large plant model' that understands plant structure across regions and species. This allows farmers to instantly adjust the LaserWeeder to their own definitions of crops and weeds.

Unlike conventional vision systems that require time-consuming retraining whenever a new crop, weed, or geography is introduced, Carbon's large plant model is pre-trained on millions of images of young plants collected worldwide. It is then personalized in the field. Using an iPad app, farmers review thumbnails from their own fields and tag a small number as 'crop' or 'weed.'

With this tool and the new foundation model, farmers can customize the autonomous weeder to 'zap' the specific plants they want removed, even if those plants are desirable in other regions or parts of the farm. The model adjusts immediately without the need to roll out new models or download software. The same global model can selectively remove weeds in a carrot field in Arizona and then move to lettuce or herbs on another farm with only minutes of configuration. This adaptability ensures quick deployment across diverse fields and crops, saving time and reducing complexity.

When the LaserWeeder destroys weeds, the nutrients from the weeds return to the soil, fertilizing the crop.

Alex Sergeev, chief technology officer at Carbon Robotics, explained the system's capabilities to The Robot Report. He described building a process that allows farmers to give examples, enabling the model to work instantaneously without retraining. The model understands differences by comparing examples to plants observed in the field at high speed, which is a fundamentally different principle from traditional methods.

Sergeev also clarified that instead of simply outputting confidence scores for weed or crop classification, the model provides a way to compare each plant against the farmer's defined crop and weed examples. This comparison must happen very quickly to support real-time decisions, ensuring accuracy and efficiency.

To train the plant foundation model, Carbon Robotics partnered with data annotation company iMerit. iMerit delivered the training resources and workflow to handle millions of plant images, a critical step for the model's effectiveness.

Alex Sergeev recounted the early stages, noting that the company initially labeled images internally but found it unscalable. In 2020, they sought a label partner and found iMerit. Sergeev estimated that Carbon had labeled about 2,000 images, while iMerit has since annotated around one million.

Sergeev also highlighted the collaborative development of a custom labeling tool designed for Carbon's specific style. The tool was provided to iMerit and could be updated based on innovation from the learning team and feedback from iMerit. This synergy allowed for rapid tool improvement and effective use.

In a related development, Sudeep George, chief technology officer at iMerit, participated in The Robot Report's webinar on 'Physical AI and Robotics.'

Carbon Robotics is also advancing tractor automation with the release of Carbon ATK, an autonomy kit for retrofitting John Deere 6R, 8R, 8RX, and 8RT (2019+) tractors without permanent modifications.

The kit includes cameras and sensors for perception and positioning, enabling obstacle avoidance, path planning, tracking, and mission definition. When integrated with Carbon's smart implements like the LaserWeeder, the tractor can operate fully autonomously without a driver onboard.

Beyond weeding, the autonomous tractor can handle tillage, spraying, and harvesting, offering a versatile solution for modern farming operations. This positions Carbon Robotics as a leader in agricultural automation.