Carbon Robotics has built an AI model that detects and recognizes plants

What to weed in the field is determined through the farmer’s eyes – and now, increasingly, by a new AI model from Carbon Robotics.

Seattle-based Carbon Robotics, which builds the LaserWeeder — a fleet of robots that use lasers to kill weeds — announced a new AI model, the large plant model (LPM), on Monday. This model recognizes plant species instantly and allows farmers to target new weeds without having to retrain the robots.

The LPM was trained on more than 150 million images and data points collected by the company’s machines across more than 100 farms in 15 countries where the robots currently operate. The model now powers Carbon AI, the artificial intelligence system that acts as the brains inside the company’s autonomous weed-killing robots.

Paul Mikesell, founder and CEO of Carbon Robotics, told TechCrunch that before LPM, every time a new type of weed showed up on the farm — or even the same type of weed in different soil or with a slightly different appearance — the company would have to create new data labels to retrain its machines to recognize the plant.

This process took about 24 hours each time, Mikesell said. Now, an LPM can instantly recognize a new weed species, even if it has never seen it before.

“A farmer can live in real time and say, ‘Hey, this is a new weed,’” Mixell said. “I want you to kill her,” and that was something that had never happened before.” “There is no new classification or retraining because the macrophyte model understands, at a much deeper level, what it is looking at and what type of plant it is.”

Mikesell said the company, which was founded in 2018, began developing this model shortly after it began shipping its first devices in 2022. Mikesell has years of experience building these types of neural networks from previous roles at Uber and working on Meta’s Oculus virtual reality headsets.

This new model will arrive on the company’s existing systems through a software update. From there, farmers can tell the machine what to kill and what to protect by selecting images collected by the machine in the robot’s user interface.

Carbon Robotics has raised more than $185 million in venture capital from backers including Nvidia NVentures, Bond, Anthos Capital and others. Now, the company will look to continue fine-tuning the model as the devices continue to be fed new LPM data.

“We now have more than 150 million plants labeled in our training set,” Mikesell said. “We now have enough data that we can look at any image and determine what kind of plant it is, what type it is, what it’s related to, what its structure is like, without ever having seen that particular plant before, because we have so much data going into the neural network.”