AI Model Identifies Plants with Precision
▼ Summary
– Carbon Robotics has developed a new AI model called the Large Plant Model (LPM) that instantly recognizes plant species for its autonomous weed-killing robots.
– The LPM is trained on over 150 million photos and data points collected from the company’s robots operating on more than 100 farms across 15 countries.
– Previously, the system required a 24-hour retraining process for new or visually different weeds, but the LPM can now learn to identify new weeds instantly without retraining.
– Farmers can now instruct the robots in real-time by selecting photos in the interface to define what plants to kill or protect, enabled by a software update.
– The company, which has raised over $185 million, will continue to refine the model as its machines feed new data into the system.
The distinction between a valuable crop and an unwanted weed has long relied on a farmer’s judgment, but a new artificial intelligence system is bringing unprecedented precision to this critical task. Carbon Robotics, the company behind the laser-equipped LaserWeeder, has unveiled its Large Plant Model (LPM), an AI capable of instantly identifying plant species. This advancement allows farming operations to target new weed varieties without the lengthy process of retraining their robotic systems, marking a significant leap in agricultural automation.
This sophisticated model is the product of extensive real-world data, trained on more than 150 million images and data points gathered from the company’s machines operating on over 100 farms across 15 countries. The LPM now serves as the core intelligence, or Carbon AI, within the firm’s autonomous weed-killing robots. Previously, the system required manual intervention whenever an unfamiliar weed appeared, even if it was a known species presenting with a slightly different look due to soil conditions or growth stage. Each instance necessitated creating new data labels and a retraining cycle that took approximately a full day to complete.
According to company founder and CEO Paul Mikesell, that cumbersome delay is now eliminated. The LPM can learn to recognize a new plant instantly, even upon first encounter. “A farmer can now operate in real time, point out a new weed, and instruct the system to eliminate it,” Mikesell explained. “This capability was previously impossible. There’s no need for new labeling or retraining because the Large Plant Model understands plant biology at a much deeper, fundamental level.”
Development of the LPM began shortly after Carbon Robotics started shipping its first machines in 2022, building on Mikesell’s prior experience constructing neural networks for companies like Uber and on Meta’s Oculus projects. The model will be deployed to existing robotic fleets via a straightforward software update. Once installed, farmers can simply review images captured by the robots in the user interface, selecting which plants to protect and which to eradicate.
The company, which has secured over $185 million in venture funding from investors including Nvidia NVentures and Bond, plans to continually refine the LPM. As the robots traverse fields, they will feed a constant stream of new visual data back into the system, further enhancing its accuracy and knowledge base. “With 150 million labeled plants in our training set, we’ve reached a point where the system can analyze any plant image,” Mikesell stated. “It can determine the species, its relatives, and its structure, even without a prior specific example, due to the immense volume of data processed by the neural network.” This continuous learning loop promises to make robotic weed control increasingly adaptive and effective for modern agriculture.
(Source: TechCrunch)
