US-based agricultural robotics company Carbon Robotics has launched the Large Plant Model (LPM), an artificial intelligence model developed for plant detection and identification in agricultural environments. Trained on crops, weeds, soil types, climates, and growth stages across multiple geographies, the LPM provides a broad base for plant recognition and decision-making.
According to Carbon Robotics, the model has been trained on 150 million labeled plant images spanning crops, weeds, climates, and growth stages worldwide, allowing farmers to deploy laser weeding across any field or crop within minutes. This Large Plant Model underpins Carbon AI, the decision-making engine that powers all Carbon Robotics products, including the LaserWeeder and the Carbon ATK (Autonomous Tractor Kit). Carbon AI analyzes vast volumes of plant and field data in real time to identify and eliminate plants, navigate field conditions, and adapt dynamically to crop variability.
When our robots can understand any plant in any field immediately and adapt behavior in real-time, farmers immediately get maximum value from the machines. The Large Plant Model provides farmers with the most advanced AI technology to maximize the weeding quality of LaserWeeder in their unique environments.
As Carbon Robotics’ LaserWeeder systems operate in fields, the platform continuously incorporates field-level data, contributing to iterative model updates and incremental performance improvements across deployed units. The company described the platform as world’s first Large Plant Model, designed to support plant level recognition across a range of farming conditions.
The company is showcasing the LPM and related technologies at Fruit Logistica in Berlin and World Ag Expo in California, as growers continue to seek solutions that reduce labor dependence, limit herbicide use, and improve crop consistency and yield.
Plant Profiles
Alongside the LPM, Carbon Robotics has introduced Plant Profiles, a new feature being rolled out across LaserWeeder systems that illustrates how the model is applied in field operations. The feature allows operators to adjust the underlying LPM to specific crops, weeds and field conditions.
Also read: Leaf Agriculture Launches LeafLake Data Platform for Agronomic Analysis
By selecting two to three reference images in the iPad Operator App, the system updates its operating parameters in near real time, enabling the LaserWeeder to respond to local field variations. As per Carbon Robotics, other AI based approaches typically require larger datasets and longer development timelines before model changes can be implemented.
We use plant profiles in our Vidalia Onion seed beds, transplants, and direct seeded onions. This has been a game changer for us and the simple, user-friendly platform allows our operators to maximize LaserWeeder performance in real-time in the field.
The company stated that usability was a key consideration in developing the system, noting that advanced AI has limited value if it requires specialist training to operate. To address this, Carbon Robotics developed an interface intended to allow farm operators to adjust a foundation model directly in the field without technical expertise.
Foundational Recognition Layer
The earlier model for agricultural robotics focused on building and training systems for individual fields, an approach that has struggled to scale. A different direction is now emerging, centered on machines that can interpret plant level information, improve through repeated deployment, and be adjusted by farmers as conditions change.
Control over a foundational plant recognition layer could become a defining element of future agricultural infrastructure. As robotics, tractors, and aerial systems increasingly rely on automated crop and weed identification, plant recognition is likely to become a shared requirement across equipment categories rather than a standalone capability.
Although the Large Plant Model was only recently introduced by Carbon Robotics, it signals a shift in how agricultural AI systems are being designed, with an emphasis on adaptability and continuous learning rather than static, field specific training. If this approach holds, its influence is expected to expand as the model incorporates new data through ongoing use.