Aigen: solar-powered weeding robots trained with NVIDIA Cosmos and run on Jetson

Weeds that no longer respond to herbicides are spreading; NVIDIA's account cites 275 resistant species documented worldwide. Removing them by hand does not scale, and spraying more chemicals brings drift and runoff. Aigen's answer is a fleet of autonomous, solar-powered robots that cut weeds out mechanically, which only works if the robot can tell crop from weed reliably in changing light, soil and growth stages.

Training that perception the usual way is slow. Aigen's CTO puts the conventional cycle of collecting, labeling and training at three to four years, and fields look different from one week to the next, so real data quickly goes stale. On board, the robot has to run perception, navigation and arm control on solar power.¹

Synthetic field data with Cosmos

According to NVIDIA, Aigen built Alchemy, a world model that generates variations of crops, weeds, soil, lighting and depth, and Cosmos 3 is its main base model. Earlier, Aigen engineers post-trained Cosmos Transfer 2.5 on about 3,000 short RGB and depth clips captured by the fleet in soybean, cotton and tomato fields, using depth as the control signal; their recipe in NVIDIA's Cosmos Cookbook says the post-trained model now feeds Aigen's perception training pipeline. NVIDIA adds that Aigen post-trained a policy on action-conditioned driving data and used it to produce about one million hours of action video within months. Aigen launched Alchemy publicly in August 2026 without naming NVIDIA in its release.

On-robot compute with Jetson

NVIDIA reports that the latest Element robots carry three Jetson Orin NX modules, which handle sensing, driving and the weeding arms and took the place of five earlier system-on-chip devices, freeing room for a four-arm layout. Aigen launched the four-arm Element Gen 2 X4 in September 2026; its release describes custom edge AI but does not name the hardware. A March 2026 AWS architecture post co-written by Aigen engineers says the edge models execute on the robot's neural processing unit, rated at 2.3 TOPS, which suggests the robots have used different compute over time; the sources do not say which units in the field carry Jetson.

Fleet status

In September 2026 Aigen said it had built more than 100 robots, logged over 15,000 autonomous field hours, deployed with Fortune 500 companies, and made fleets available to growers for the 2027 season.¹²³⁴⁵

What others can learn: an agricultural robot maker can capture its own field video and post-train an open world model to generate labeled training images for new crops, rather than waiting seasons for real data, and consolidating several onboard computers can free power and space for more working tools. Limits: the 1 percent real data and 30 percent utilization figures come from NVIDIA without published accuracy or test conditions, the sources disagree on which onboard processor the robots use, and no source ties a weed control or yield result to the NVIDIA components. Farms with herbicide programs that still work, or with small plots, may be served by cultivators with camera guidance or by targeted spraying services without owning robots.