I had the opportunity to interview Paul Mikesell, CEO of Carbon Robotics. Carbon Robotics visually identifies crops and zaps weeds with lasers. The interesting thing is you will learn about how they manage to visibly monitor a variety of crops, terrain issues, and of course weeds. Carbon Robotics already has a global presence and is well positioned for continued growth.
A little about Paul: He is the founder and CEO of Carbon Robotics, with deep experience founding and building successful technology startups. Paul co-founded Isilon Systems, a distributed storage company, in 2001. Isilon went public in 2006 and was acquired by EMC for $2.25 billion in 2010. In 2006, Paul co-founded Clustrix (News - Alert), a distributed database startup that was acquired by MariaDB in 2018. Prior to Carbon, Paul served as Director of Infrastructure Engineering at Uber, where he grew the team and opened the company’s engineering office in Seattle, later focusing on deep learning and computer vision.
Carl Ford (News - Alert): First of all, congratulations; based on the sales success and the investors of Carbon Robotics, you are well on your way to becoming a dominant force in the industrialized farming market. If you can briefly explain the idea of LaserWeeder, and how you determined what size machines you were going to build?
Paul Mikesell: Our mission, from the beginning, has been to develop practical, high-performance tools that solve real, on-the-ground problems for farmers, such as persistent labor challenges, herbicide-resistant weeds, and rising input costs. LaserWeeder came out of a pressing need to reduce herbicide use and labor dependency in weeding, two issues that severely limit profitability and sustainability in modern agriculture.
We saw an opportunity for computer vision, AI/deep learning, robotics and high-powered lasers to converge into a single, elegant solution. Lasers offered precision, speed, and a chemical-free method of eliminating weeds with zero soil disruption, making them much cleaner and a more scalable technology than legacy alternatives on the market. Early trials confirmed that LaserWeeder met a real demand among growers looking for chemical-free, scalable alternatives.
When it came to determining the size of the machine, we looked directly at what large-scale commercial farmers were using and where the pain points existed. These operations already run implements across vast acreage, so we decided early on that LaserWeeder needed to integrate seamlessly into that ecosystem. But, we also recognized that not every farm is the same. That’s why, with LaserWeeder G2 (News - Alert), we’ve expanded the product line to include multiple size configurations, with width options ranging from 2 to 12 meters, allowing us to serve a broader range of farm sizes and geographies. Whether managing a small specialty farm or thousands of acres, growers can now choose from a range of LaserWeeder models designed to meet their operational scale.
CF: You have demonstrated a diversity of crops that you support, but I have always understood that different crops breed different problems. Does this mean you have developed your AI capability based on one large language set? Is the AI using a visual language set? Do geographic regions (i.e., rocky terrain) come into play here?
PM: We don’t use a traditional large language model like you would for text-based AI systems. Instead, our AI is built on a custom-trained visual perception model. It’s essentially a neural network explicitly trained to distinguish between crops and weeds in a wide range of conditions. So, it’s not a "language" model in the classic sense, but rather a visual intelligence system that can process high-resolution imagery in real time and make millisecond-level decisions about what to target and what to preserve.
We trained the AI using millions of annotated plant images. These datasets are incredibly diverse with different crops, weeds, growth stages, lighting conditions and soil types. That variety is what gives our system its strength in the field. It’s not about using a single language set, but about building a flexible, evolving visual recognition system that learns from the field and improves with more exposure.
Geographic and terrain factors are also critical. The system has to perform consistently whether you’re in flat, loamy fields in California or dealing with rocky, uneven terrain in Colorado or the Pacific Northwest. Our models are trained and validated across these diverse environments. In fact, one of the advantages of our platform is that it’s designed to improve continuously. As the system encounters new scenarios, new weed types, different lighting, and terrain variations, it can feed that data back into the model for retraining and refinement, ensuring even better performance over time.
CF: Speaking of AI and your connectivity, you are designed for large farms, which probably have little to no coverage of cellular networks. How do you manage connectivity?
PM: LaserWeeder is designed to operate effectively even in areas with limited or no cellular coverage. The system is equipped with onboard computing capabilities, including high-resolution cameras and NVIDIA (News - Alert) GPUs, allowing it to process data and make decisions locally in real time, without relying on external communication networks.
LaserWeeder also utilizes a combination of connectivity solutions. It features a GPS/LTE (News - Alert) antenna for precise positioning and is equipped with satellite internet connectivity to receive and send real-time updates. Additionally, operators can monitor and adjust key settings in real time from the cab via the iPad operator app.
This combination of onboard processing and multiple flexible connectivity options ensures that LaserWeeder can function reliably across diverse farming environments, regardless of cellular network availability.
CF: Looking at the size of the machine and your success in selling to Europe and Australia, I am assuming that some assembly has to be done once the tractor has arrived. Does that add any difficulties?
PM: We’ve engineered LaserWeeder to be straightforward to assemble once it arrives on-site by trained Carbon Robotics personnel located in-region, whether it’s in the U.S., Canada, Europe, or Australia. Given the size of the equipment and the logistics involved in global shipping, a weeding module-based design was intentional from the start. All LaserWeeders are designed to attach to tractors’ standard rear CAT 3 3-point hitch.
CF: How do farmers pay for the service, and does it work like Caterpillar renting it out?
PM: Our customers purchase LaserWeeder outright. It’s a capital equipment purchase, like a tractor or a sprayer. Most farmers realize a payback period on their investment in one to three years. This is driven through significant weed control cost savings (labor, herbicide, etc.) and an increase in crop yields, quality, and consistency.
CF: How often are there updates in the AI software and is the software primarily in the cloud or are the units more like an edge compute system?
PM: We release significant software updates to the entire worldwide fleet on a periodic basis. Also, we have the ability to create and download new crop models to LaserWeeders in as little as 24-48 hours, which is unparalleled in the industry. We push updates regularly to improve accuracy, expand weed identification, and enhance overall performance, so growers benefit not just from the initial technology but from a platform that keeps getting better and better over time. LaserWeeders in the field each operate independently but maintain a bi-directional data and communication link with Carbon Robotics.
CF: Zapping a weed with lasers sounds very cool, but how do you do this while moving? Are there multiple lasers involved? What if a particular area is infested with a foreign growth? Does the tractor stop when it needs to deal with a patch of weeds?
PM: LaserWeeder uses multiple high-powered lasers – up to 30 – and advanced precision targeting systems to eliminate weeds while the machine moves at optimal speeds. We provide operators with a velocity estimator in the operator app so they can adjust tractor speed up or down as weed pressure density and size conditions change. This ensures the optimal tradeoff between speed/acres covered and weed elimination efficiency.
CF: What does the future hold for Carbon Robotics? What else will you manage, like pest control or water preservation?
PM: We’re focused on building technologies that deliver measurable gains in productivity, profitability, and long-term sustainability. Our goal is to empower farmers with innovative technologies that not only improve their productivity but also increase profitability and sustainability in agriculture, ensuring a resilient future for farming.
CF: What are plans are there for offering services to smaller farms?
PM: We’re committed to supporting smaller farms. That’s why our LaserWeeder G2 comes in multiple sizes, starting as small as two meters wide, making it accessible for specialty crop growers with as little as 200 harvested acres. While our initial focus was large-scale farms for fast adoption, we see tremendous value in expanding access to smaller growers who face the same pressures around labor and input costs. Long term, our goal is to make laser weeding a pervasive standard tool used across farms of all sizes.
CF: Given recent government policies about immigration, has that increased interest in Carbon Robotics? What kind of impact do you have on manpower?
PM: Recent immigration policies have continued to put pressure on the lack of availability and rising cost of labor on farms, so more and more growers are focused on finding new ways to automate weed control. We are seeing tremendous worldwide demand for Carbon Robotics' solutions with more than 100 growers in 13 countries. Carbon Robotics' LaserWeeder G2 and AutoTractor offer AI-driven weed control and tractor operations, reducing reliance on human labor. These technologies operate 24/7, providing consistent performance even in labor-constrained environments. By integrating such automation, farms can maintain productivity and mitigate the impacts of workforce shortages.
CF: I know you are a no-nonsense company, but I have to ask, do your tractors make “Farmer’s turns?”
PM: While we don’t build tractors ourselves, our AutoTractor tractor autonomy solution is designed to make efficient row-end turns. Turnaround time is a metric that we measure and focus on optimizing. Headland space and obstacles vary considerably across regions and farms, so I think it’s fair to say that some turnarounds are just prettier than others.