Aigen uses synthetic fields to train weeding robots for new crops

Aigen says its synthetic fields help its weeding robots learn new crops in about a week, where collecting and labeling field images once took years. The Redmond, Washington, company launched its Alchemy training system in August. In an October 23 interview with GeekWire, co-founder and chief technology officer Richard Wurden said the faster training has let Aigen sell robots to growers of crops it previously could not serve.

The big change

• What changed: Aigen is using generated, labeled images to vary crops, weeds and field conditions without waiting for each growing season to supply examples. Its new-crop training time and sales account come from the company; the August system launch is the earlier event.
• Why it matters: A robot that mistakes a crop for a weed can damage the plant it was meant to protect. Faster image production could widen the crops Aigen can support, while the field remains the test of whether those generated examples prepare the machine for the plants it encounters.
• What to watch: Aigen says it is taking orders for the 2027 season. Crop-by-crop field results, including weed removal and crop damage under varied conditions, would show how much of the training gain carries over to farms. The public materials reviewed here do not provide that comparison.

From field footage to labeled training scenes

Alchemy draws on data gathered by Aigen's own robots and generates variations in plants, soil and light. The company's Alchemy page describes controlling conditions such as a weed's growth stage, partial obstruction and camera position. Because the system generates the scenes, Aigen says it can supply labels for the plants it depicts without a person marking each image. Its Element robots then use onboard cameras and AI to distinguish crops from weeds and direct arms to remove weeds mechanically.

A crop and a weed may overlap in a camera view; their appearance changes with growth, daylight and soil. Aigen can generate examples of those combinations without waiting for another planting season. Its August Alchemy announcement says some Element training sets are up to 100% synthetic. Its current product page says production models use 99% to 100% synthetic data. Those percentages describe the training mix; they give no measured rate of correct weed removal.

NVIDIA's Aigen case study gives more detail about one part of the pipeline: Aigen used field camera and depth footage to adapt a video-generation model, then used synthetic agricultural scenes for perception training. NVIDIA and Aigen are technology partners, so the case study supplies a technical account rather than independent field validation. Aigen has not published, in the materials reviewed for this article, a crop-by-crop comparison of synthetic and conventionally trained robots under the same field conditions.

A shorter training cycle meets a longer field test

Wurden told GeekWire that collecting and labeling real images for a new crop had taken three to four years and now takes about a week. He described sales at the FIRA USA farm-robotics show to growers Aigen had not previously served, and said robots for the 2027 season were nearly sold out. Those are the company's accounts to the reporter. They do not establish how many new-crop robots have completed a season or how their weed-removal and crop-protection rates compare with earlier training methods.

Aigen's September Element Gen 2 X4 announcement says the four-arm robot is available for the 2027 season and reports more than 15,000 autonomous field hours across its robot generations. Field hours show operating exposure, but the announcement does not break out the new crops identified in the October interview or connect those hours to a measured Alchemy training advantage.

Synthetic images can multiply the examples available to train perception. The robot must still identify plants in changing real conditions and place its weeding arms without harming the crop. Aigen's reported new-crop sales give that training question immediate commercial relevance. Comparable field results would show how much of the claimed speed gain carries over to the crops now being sold.

Sources & further reading

• GeekWire's October 23 interview with Richard Wurden reports the new-crop training-time and sales claims, the FIRA USA setting, and Wurden's account of customers. It is original reporting based on a company interview, not an independent crop-performance trial.
• Aigen's August 3 Alchemy announcement describes the launch and the company's synthetic-data and deployment claims. The Alchemy product page explains the scene variables and current training-mix claims. Both are company descriptions; neither supplies comparative field error rates.
• NVIDIA's Aigen case study describes the partner's model-adaptation and synthetic-training pipeline. NVIDIA supplies technology used by Aigen, and its account does not independently quantify Alchemy's field-transfer advantage.
• Aigen's September 16 Element Gen 2 X4 announcement gives the robot's release stage and company-wide field-hours claim. It does not isolate results for the new crops discussed in the October interview.