Quiet Breakout: The Fields Are Getting 240-Watt Lasers and 36 Cameras

A LaserWeeder G2 is a forty-foot trailer dragged behind a tractor through an onion field in Washington. Mounted underneath are 36 high-resolution cameras, a rack of Nvidia GPUs, and 30 fiber-coupled CO2 lasers rated at 240 watts apiece. The cameras classify every plant that passes under the rig. The GPUs decide which pixels are an onion and which are a weed. The lasers fire. A weed cooks from the inside in something like 50 milliseconds. The rig can do roughly 600,000 of these per hour, according to Carbon Robotics, the Seattle startup that builds it.

Nobody in the tech press writes about this much. Agriculture Dive and AgFunderNews and Red River Farm Network do. Reuters and Wired and TechCrunch mostly do not. That is the story.

In the patent record, the shape of the story is unmistakable. The US Patent Office granted Blue River Technology, a Sunnyvale startup John Deere bought in 2017 for $305 million according to the original AgFunderNews announcement, five grants on computer-vision weed control from 2015 through 2017. Since January 2020 it has received 57. Carbon Robotics, founded in 2018 by former Isilon CEO Paul Mikesell, had zero US grants before 2020. It now has seven, including a 2023 grant (US11602143) that claims a two-camera architecture the patent itself calls “prediction” and “targeting.” A wide-field camera hands off coordinates to a narrow-field camera that closes the loop on a steered mirror, which then fires the beam. The language comes straight from missile guidance. The target is a pigweed seedling.

Why the silicon arrived

The International Herbicide-Resistant Weed Database, run out of the University of Queensland and Iowa State, records 274 weed species with confirmed resistance to at least one herbicide, across 546 unique species-by-mode-of-action cases as of late April 2026. Palmer amaranth, a pigweed that can grow three inches a day and outcompete soybeans, has populations confirmed to have evolved resistance to as many as six distinct sites of action. Glyphosate-resistant Palmer amaranth was first identified in Georgia cotton in 2005, per USDA-ARS. It has since been documented in counties in New York where glyphosate was the primary in-crop herbicide, according to a Weed Science paper. The chemistry is running out of novelty.

Research output tells the same story from the other side. In the OpenAlex scientific literature corpus, papers combining the phrases “weed detection” and “deep learning” have gone from two in 2016 to 235 in 2025. A single 2018 paper in Remote Sensing on UAV weed detection with unsupervised labeling now has almost 300 citations. The computer-vision community found a stuck biological problem, and pointed a GPU at it.

The companies

Strip the US patent grants since 2020 down to filings whose abstracts literally describe a machine that uses an image to decide where to spray, cut, or fire at a weed, and a small constellation appears:

That last row is the one to sit with. Climate LLC is the software arm of Bayer Crop Science, which inherited Roundup from Monsanto in the 2018 merger. Bayer Aktiengesellschaft itself kept patenting new herbicidal chemistry through 2025. At the same time, its Climate subsidiary has been filing vision-based systems that make heavy chemical application unnecessary. A company that sells the cure is patenting the thing that makes the cure less necessary. That is not a contradiction so much as a hedge, and it is the clearest signal in the data that the center of gravity in row-crop weed control is shifting from the drum of herbicide to the rack of GPUs bolted under the sprayer.

What the machines actually do

The coherence that holds this category together is not a keyword. It is a four-part engineering pattern that repeats across every assignee above. An imaging stack, often multi-spectral and sometimes stereo, feeds frames to a convolutional network that has been trained on hand-labeled crop and weed images. The network emits what Blue River’s US11514671 calls a “treatment map,” specifying, for every pixel of soil the machine is about to roll over, whether a weed lives there and what to do about it. A precision actuator, bolted to the same chassis and synchronized to wheel odometry, executes.

The actuator is where the designs fan out. Blue River and Deere chose nozzles: tiny solenoid-valved injectors, each aimed at a band of soil a few centimeters wide, that can pulse a jet of non-residual herbicide onto a single weed. Carbon Robotics chose thermal: 30 CO2 lasers, each coupled to a galvanometer-steered mirror, denaturing the weed’s meristem in tens of milliseconds. Earth Rover’s UK patent (US12185712) describes a different thermal path, an array of cheap semiconductor emitters focused through a lens stack into a convergent beam, which cooks the weed with concentrated visible light rather than a single high-power laser. FarmWise went mechanical, swinging hoes at weed-sized gaps between crop plants. EcoRobotix lowered a spray arm to within millimeters of the canopy to cut drift. Each approach is a different answer to the same control system.

The numbers for the reader deciding budgets

John Deere announced in November 2025, via Red River Farm Network, that its See & Spray product had been used on five million acres that season. In field trials reported by the same outlet, operators averaged roughly a 50 percent reduction in non-residual herbicide, even in a wet Corn Belt spring that drove weed pressure up. An Iowa State University precision-spraying study, reported by Strip-Till Farmer, found $15.70 per acre in herbicide savings. Deere’s own soybean trials, run with seven universities, found a two-bushel-per-acre yield bump, four bushels in some plots, which the company attributes to less unintended crop exposure to herbicide.

Carbon Robotics reports its fleet has passed 230,000 acres and prevented over 100,000 gallons of herbicide application. In April 2025 the company raised another $20 million, per GeekWire, on top of a $70 million Series D the previous October that had already taken total funding past $150 million. Taylor Farms, the largest producer of packaged salads in North America, bought the FarmWise business outright on April 4, 2025. A fresh-produce buyer vertically integrated its own fleet of robot weeders. That is what it looks like when a supply chain stops waiting for a technology to mature.

For R&D directors at the major agrochemical companies, the implication is uncomfortable. The market for herbicide volume is quietly decoupling from the market for weed control. The market for weed control is migrating to a bundle of machine vision, compute, and actuator engineering that looks a lot more like industrial robotics than like crop science. For venture investors, the interesting question is no longer whether the technology works. Deere’s five-million-acre 2025 campaign settled that. The question is whose actuator, whose imaging stack, and whose data moat ends up as the default fixture under the next generation of tractors sold into 300 million acres of US row crop.

The machine under the tractor was invented in 2017. It was granted real patents in 2021 and 2022. It took the field in 2024 and 2025. Each Signalnet piece asks the same question: when did this story actually become visible in the data? For precision weed control, the honest answer is more than five years before the tech press noticed.

Method note

Patent counts come from a local copy of the USPTO utility grant feed, roughly 9.3 million US grants, filtered to assignees whose abstracts describe image-driven weed identification and precision application between January 2020 and April 2026. The research-output trend comes from OpenAlex, an open index of about 357 million scientific works, counting papers whose title or abstract matches “weed detection” and “deep learning.” Herbicide-resistance counts come from the International Herbicide-Resistant Weed Database (weedscience.org), as of its April 23, 2026 update. Adoption and funding figures come from Red River Farm Network, GeekWire, AgFunderNews, Strip-Till Farmer, and Taylor Farms’ own April 2025 press release, attributed to the outlet named in the sentence that cites them. Assignee counts collapse variant spellings but do not roll subsidiaries up into parents. Climate LLC is reported separately from Bayer Aktiengesellschaft even though Climate is a Bayer unit. Patent filings are a leading indicator of R&D commitment, not of commercial success; several of the private companies named above may not survive another fundraising cycle.