Physical AI on the farm.
US agriculture is counted well at the top and barely at all where a machine has to work. The 2022 Census enumerates 1,900,487 farms on 880 million acres and 3.78 million tractors, and then records no engine hours, no field passes and no count of the acres on which hand weeding is actually performed. This review prices what Physical AI can take on against that background, across the estate, the labour market and the logistics layer.
Ronnie Gomez, Industrial Research Fellow · The Charlot Lab
46 years
to turn over the tractor fleet
At 2024 new-machine retail rates, for the 100-horsepower-plus class, and about 54 years for combines. Autonomy that ships only inside new iron arrives slowly.
modelled on census counts
83.2%
of the hired-labour bill, in 3.5% of farms
67,320 farms carry it. The population that labour-displacing automation addresses is small, identifiable and already known to equipment dealers.
verified (2022 Census)
32.3%
cut to the California H-2A wage floor
From an operative $19.97 to $13.52 under the 2025 interim final rule. That is the largest single input to every specialty-crop robotics business case.
verified (DOL)
0.14%
of farm production expenses
What automated weeding is worth at full national adoption across all vegetable acres excluding potatoes, about $612M a year against a $424bn expense base.
modelled
The estate, the wage floor and the fleet.
What the recurring work of farming actually costs, which farms could buy their way out of it, and why the wage floor those farms face was rewritten while the machines were still being built.
The estate is enumerated, and the work is not. The 2022 Census counts 1,900,487 farms on 880,100,848 acres, with the largest 2% of farms holding 42% of the farmland, and it counts a machine fleet of 3,784,743 tractors and 298,301 self-propelled combines. It records no engine hours, and farm tractors are not registered or licensed anywhere in the United States, so the installed base has no utilisation measure at all. At 2024 new-machine retail rates the 100-horsepower-plus tractor fleet would take about 46 years to replace and the combine fleet about 54, which is why retrofit and supervised operation, not new iron, carry the near term.
The labour bill is concentrated. Hired and contract labour is 12.24% of a $424.14bn census expense base, 23.0% of farms report any hired-labour expense at all, and 67,320 farms carry 83.2% of the national hired-farm-labour bill. That is the addressable population, and it is small enough to name. The wage floor those farms face then moved by a third: the Department of Labor's 2025 interim final rule cuts the California entry-level H-2A rate from an operative $19.97 to $13.52. USDA discontinued the Agricultural Labor Survey on 28 August 2025, so the series that would have measured the effect ended in the same season the rule changed it.
The deployed record is thinner than the marketing, and the withdrawals are better documented than the installations. Monarch Tractor ceased operations on 3 April 2026. FarmWise wound down after raising about $95.7 million, and advanced.farm and Tortuga AgTech exited within four weeks of it in March 2025. The largest verified embodied deployment in US agriculture is robotic milking, at 6% of milk in 2021, which works because the cow presents herself indoors on a prepared surface. California still requires an operator at the controls of self-propelled equipment under a rule written in 1977.
23 sourced records.
Every deployment, withdrawal and independent measurement the review tracks. The closures carry the strongest evidence in the table, because a closure is a filing and a launch is a press release.
Every deployment, withdrawal and independent measurement the review tracks, with the grade of the evidence behind each. A record graded self-published comes from a party selling the thing being measured; the independent measurements are the small minority graded verified or reported.
Autonomy, guidance and milking (7)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Robotic milking | Lely | self-published | Astronaut automatic milking robot, product line since 1995Source [33] in TR-2026-32. | Over 50,000 cumulative installations in 40+ countries; not units in service |
| Robotic milking, independent measure | USDA ERS | verified | ARMS-based adoption estimateSource [34] in TR-2026-32. | 6% of US milk in 2021, 4% in 2016; 13% of 150-499 head farms |
| Guidance autosteering | USDA ERS | verified | Operator-assisted guidance, driver in the cabSource [35] in TR-2026-32. | 52% of midsize and 70% of large-scale crop farms in 2023 |
| Autonomy retrofit, permanent crops | Bluewhite (acquired by Elbit FUSE, 27 May 2026) | self-published | Retrofit kit for existing orchard and vineyard tractorsSource [41] in TR-2026-32. | 150,000 US acres, 20+ growers, 75,000+ hours; scope figures unreconciled |
| Autonomy retrofit, row crops | Sabanto | self-published | Retrofit kits for conventional row-crop tractorsSource [42] in TR-2026-32. | About 300 units shipped, US and Australia; investors also sell the inputs applied |
| Purpose-built autonomous platform | SwarmFarm Robotics (Australia) | self-published | Small autonomous platforms carrying third-party implementsSource [43] in TR-2026-32. | 250+ robots, 10M acres claimed Nov 2025; cumulative-or-annual not stated |
| Monarch commercial record | Monarch Tractor | self-published | Company statements made while seeking a buyerSource [39] in TR-2026-32. | About 500 tractors delivered; over 70% of 2025 revenue from licensing |
Weeding and spraying (7)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Targeted spraying | Deere See and Spray | self-published | Camera-triggered nozzle control, operator in the cabSource [36] in TR-2026-32. | Over 5M acres in 2025, about 1M in 2024; 59% claimed herbicide saving |
| Targeted spraying, independent test | Iowa State | reported | Five Iowa fields, 2024Source [78] in TR-2026-32. | 76% average material saving, $15.70/acre; 43.9% to 90.6% spread |
| Precision sprayer | Ecorobotix ARA | self-published | Towed sprayer pulled by a conventional tractor with a driverSource [49] in TR-2026-32. | 1,000 units sold in 30+ countries over five years; about 4 ha/hour |
| Supervised orchard spraying | GUSS Automation (acquired by Deere, 2025) | self-published | One operator monitors several machines from a pickupSource [37] in TR-2026-32. | 250+ machines, 2.6M acres, 500,000+ autonomous hours, cumulative |
| Laser weeding | Carbon Robotics | self-published | LaserWeeder, towed, operator presentSource [44] in TR-2026-32. | 150+ machines, 100+ farms, 15 countries; no price or throughput published |
| Laser weeding economics | Braga Fresh and Triangle Farms | self-published | Grower-supplied figures in a trade-association case studySource [46] in TR-2026-32. | $267.72 and $448.73 per acre, same machine and season, two farms |
| Mechanical weeding, field trial | UC West Side REC | reported | Replicated trial, drip cantaloupe, 2022Source [77] in TR-2026-32. | Robovator saved $37/acre, about 20%; herbicide saved over $100 |
Harvest, packhouse and transport (3)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Autonomous material transport | Burro | self-published | Autonomous carts on prepared surfacesSource [50] in TR-2026-32. | 750+ robots, 800,000 hours, cumulative across all sectors |
| Robotic apple packing | Hawk Technology | self-published | Packhouse robot, 12 inspection zonesSource [56] in TR-2026-32. | 50 US installations, 250 on order; up to 2 labour units per lane |
| Robotic apple harvest | Michigan State and USDA-ARS | verified | Research prototypes, single-arm and dual-armSource [53] in TR-2026-32. | 82.4% and 65.2% success by orchard type, detection-conditional denominator; dual-arm figures rest on an arXiv preprint |
Closures, exits and litigation (6)
| System | Who | Evidence | What it is | Status |
|---|---|---|---|---|
| Purpose-built electric tractor | Monarch Tractor | verified | MK-V driver-optional electric tractorSource [38] in TR-2026-32. | Ceased operations 3 April 2026; IP acquired by Caterpillar, terms undisclosed |
| Monarch dealer litigation | Burks Tractor and two others | reported | Federal complaints from September 2025Source [40] in TR-2026-32. | Ten tractors, about $800,000 financed, alleged not to operate autonomously |
| Field robotics, long-running vendor | Naio Technologies | reported | Small autonomous weeding platforms since 2011Source [48] in TR-2026-32. | 350+ cumulative; judicial recovery June 2025, relaunched with EUR 6.4M |
| Commercial harvest robotics | Abundant, advanced.farm, Tortuga | reported | Apple and strawberry harvesting venturesSource [51] in TR-2026-32. | All three exited between June 2021 and March 2025 |
| Automated weeding service | FarmWise (acquired by Taylor Farms, 4 April 2025) | reported | Weeding-as-a-service, Vulcan machinesSource [47] in TR-2026-32. | Wound down after about $95.7M raised; buyer was investor and customer |
| Field robotics closures | Small Robot Company, Iron Ox | reported | UK weeding robots; US indoor growingSource [52] in TR-2026-32. | Liquidation Feb 2024 after $13.18M; Iron Ox closed after $102M |
The number the whole sector is priced against.
The review nominates a measurable gap a single operator could close in one season.
Every automation savings claim in specialty crops is denominated in the cost of a hand-weeding pass, and nobody publishes how many passes an acre actually receives, on which crops, in which regions. NASS publishes acres by crop and by organic status, but not by weed-control practice, so the addressable acreage for automated weeding cannot be derived from any public series. One grower, one season and a pass log would produce the denominator the sector prices itself against. The same operator could publish machine utilisation and downtime alongside it, which no vendor currently does: the fullest public unit economics in the field come from two farms in one valley, and they disagree with each other by a factor of 1.7 on the same machine in the same season.