The distinction between what constitutes a weed and what does not has traditionally relied on the farmer's knowledge. However, a groundbreaking AI model from Carbon Robotics is now enhancing this process.
Based in Seattle, Carbon Robotics, known for its innovative LaserWeeder technology that utilizes lasers to eliminate weeds, has introduced the Large Plant Model (LPM). This advanced AI model enables immediate identification of plant species, allowing farmers to address new weed challenges without the need for extensive retraining of their robotic systems.
The LPM has been developed using a vast dataset of over 150 million images and data points gathered from the company's robotic operations across more than 100 farms in 15 countries. This model is integral to Carbon AI, the intelligent system that powers the company's autonomous weed management robots.
According to Paul Mikesell, the founder and CEO of Carbon Robotics, the previous method required creating new data labels whenever a novel weed appeared on a farm, which was a time-consuming process taking approximately 24 hours. With the LPM, farmers can now teach the system to recognize a new weed in real time, even if it has never encountered that specific plant before.
"Farmers can interact with the system instantly, saying, 'This is a new weed; I want it removed,' which is a significant advancement," Mikesell explained. "There's no need for new labeling or retraining, as the LPM comprehends the plant's characteristics at a much deeper level."
Since its inception in 2018, Carbon Robotics has been refining this model, which began development shortly after the launch of its first machines in 2022. Mikesell brings years of expertise in neural networks from his previous roles at Uber and in developing Meta's Oculus virtual reality systems.
This innovative model will be integrated into the existing systems through a software update. Farmers will then be able to instruct the robots on which plants to eliminate and which to preserve, utilizing images collected by the robots within their user interface.
Having secured over $185 million in venture capital from notable investors such as Nvidia NVentures, Bond, and Anthos Capital, Carbon Robotics aims to continuously enhance the LPM as the robots gather more data.
"Our training set now includes over 150 million labeled plants," Mikesell stated. "With this extensive data, we can analyze any image to determine the plant's species and characteristics, even if it's a plant we've never encountered before, thanks to the depth of our neural network."
