01
Visual Foundation Models
Off-the-shelf vision models don't understand a vine. We fine-tune our own, purpose-built for agricultural geometry — running zero-shot across cultivars and lighting on edge compute (Jetson Orin).
12AM Robotics · Bengaluru | California
We are building
the foundational
model for agriculture.
12AM Robotics builds visual foundation models and autonomous field robots, purpose-built for specialty crops — starting with grapevines in Maharashtra.
Pruning 200+ acres of grapevines in Maharashtra, IndiaWinter 2026 deployment
Mission
Build the foundational model
for agriculture.
We are training AI to see, understand, and tend to every plant on Earth. The first crop is grapes. The first task is winter pruning. The endgame is every specialty crop, every season — autonomously.
Foundation models finally understand biological variability. Edge compute is cheap enough to put in a field robot. Synthetic data has closed the sim-to-real gap. The pieces have arrived together — once. We are picking them up.
The Problem
Specialty agriculture is losing the people who hold it together.
Wine grapes can't be wheat. They can't be combined, sprayed flat, or harvested in a single pass. They need hands — pruning, thinning, training — on millions of vines, in narrow seasonal windows. Those hands are vanishing.
We are rebuilding them in silicon.
40%
of fruit & vine production cost is manual labor
$500B+
global TAM for manual agricultural labor spend
+25%
surge in seasonal wages over the last 36 months
8 wks
annual pruning window. Miss it and the season degrades.
Sources: USDA, FAO, ag-econ field studies, 2024–2026
The Robot
An autonomous vineyard worker.
Subscribed, not bought.
Our ground robot is a perception platform with cutting arms. It moves vine-by-vine through the dormant winter rows, identifies every cane and bud, and executes an agronomically-correct pruning plan — one cut at a time.
Growers don't buy it. They subscribe per acre, the way they pay for irrigation or crop insurance. Predictable OpEx replaces volatile labor.
And while it prunes, it scans. Every plant on the farm becomes a 3D record that compounds, season over season, into the dataset that trains everything we build next.
Unit · v0Pilot-ready
- Operating window
- 20+ hrs / day
- Cost vs. seasonal labor
- Up to 40% lower per acre
- Hardware payback
- <18 months (RaaS)
- Mapping resolution
- Every plant, every season, in 3D
- Edge compute
- NVIDIA Jetson Orin
- Deployment model
- Subscribed, not bought
First commercial deployment — winter 2026, Maharashtra, India.
The Technology
A thousand seasons in simulation — before the next one begins.
Most AgTech founders think agriculture is too messy to simulate. We think the reality gap closed two years ago, and the industry is severely underestimating it. We bet the company on it.
02
Simulation-first development
We train and validate inside a high-fidelity digital twin (Isaac Sim) with deformable vines and physically-credible cuts. 10,000 pruning cycles overnight that would take a decade in the field.
03
The data flywheel
Every cut maps a vine. Every vine refines the model. Winter pruning unlocks summer harvesting on the same plants — and grapes unlock the next crop. The dataset compounds.
The data flywheel
01Robot prunes a vine
02Captures 3D scan
03Trains foundation model
04Unlocks next crop
Simulation is not a nice-to-have — it is the cheapest place to fail. We retire perception, motor-policy, and configuration risk in silicon before a single screw is turned in the field.
The Thesis
One model — any crop.
The foundation model we train on a grapevine doesn't stop at grapes. Specialty crops share more structure than the industry assumes — and a model that learns one, learns the next faster.
Traction
We don't pitch farmers.
We've already onboarded them.
Strategic partnerships with commercial cultivators in Western Maharashtra de-risk our first deployment — and give us a real-world testbed that competitors can't access.
- NowMaharashtra pilot, 200+ acresCommercial grapevine cultivators onboarded as deployment partners. Winter 2026 field season.
- Phase 1Sim-validated end-to-end pruneFull autonomous pruning demonstrated in high-fidelity simulation. Targeted at ARTPARK / IISc residency.
- Phase 2Robot in the row, human-in-the-loopSim-trained stack on real hardware in commercial vineyards. Every operator override feeds the flywheel.
- Phase 3Full autonomy + summer harvestingSame plants, same map. Pruning model becomes harvesting model. Then: next crop.
Origin
Why 12AM.
The name is honest. We started this company on the night shift — after day jobs, after dinners, after everyone else had logged off. Midnight was the only hour quiet enough to think about something this big.
The clock stuck. Even now, the hardest problems get cracked open somewhere around 12 AM, when the room is finally quiet and the vineyard is finally still.
We're still keeping that hour.
Founder
In stealth.
Robotics engineers with farming roots and years in simulation-driven robotics and sim-to-real transfer. We'll say more when we leave stealth.
“Simulation is not a nice-to-have — it is the cheapest place to fail.”
One design rule for this company: farmer-first hardware, not fragile lab tech. If it can't survive a season in a working vineyard, it doesn't ship.
Owl Call
Talk to us.
We're awake.
Three lanes, one inbox. Pick the one that fits.
Growers
Bring it to your vineyard.Request a pilot →
If you manage 50+ acres of grapes or specialty fruit and labor is your scarcest input, we want to talk.
InvestorsBack the foundation.Request the deck →
Currently in conversations for our first institutional round. The deck and full data room are available on request.
BuildersBuild this with us.Join us →
Hiring robotics, perception, ML, and field-engineering talent. If sim-to-real is the most interesting problem in robotics to you, send us anything you've built.