Verdant OS gets smarter because every farm in the network contributes labelled plant-state data back to a shared model. The research below is how that brain is built — and most of it is open.
Foundation models for plant tissue, fruit-set, pest detection. Owns VerdantBench and our internal labelling pipeline.
Multi-objective RL for climate, irrigation and lighting under noisy sensors. Maintains the simulator that ships with every OS release.
Cultivar selection, seed-line breeding, post-harvest physiology. Curates the public recipe library.
Releases the network's labelled image and telemetry data on a 6-month delay under CC-BY-NC. 12.4M images shipped to date.
We show that a small vision transformer, trained on 1.1B labeled plant-state pairs from the Verdant network, predicts SPAD chlorophyll readings to within ±2.1 μg/cm² — outperforming destructive lab assay reference at a fraction of the cost.
Twelve months of online experiments demonstrate that a federated Bayesian dosing policy reduces nitrogen overshoot by 41% while improving harvest weight uniformity across the network.
We release 12.4M labelled images across 47 crops, 9 disease classes, and 14 nutrient states — with a held-out split spanning 80 commercial farms — as a public benchmark for the next generation of plant-vision models.
A first-of-its-kind LCA across 11 countries places per-kg embodied water at 4.2 L and CO₂e at 0.31 kg for leafy greens grown on the Verdant network, including all upstream energy.
We describe the off-policy actor-critic that runs in production behind Verdant OS climate control, and how a 36-hour reward horizon was made tractable with a learned dynamics model.
We host 6–12 visiting researchers a year across the three labs. Stipend, housing, full network data access, and a real production farm to run experiments on.
