Carbon Robotics LaserWeeder G2: the perception loop that will decide whether the machine can deliver on its promise

Carbon Robotics has pushed its LaserWeeder to a second generation, and the practical question for a weeding implement is whether the G2's perception loop can keep up with field conditions. The G2 uses onboard cameras, an AI-based plant-classification model and laser actuation that targets weeds as they emerge, applying no chemical and disturbing no soil. The company markets the machine as a replacement for hand-weeding labour and herbicide application.

For a commercial farm the operational question is not whether the announcement is real but how the image, inference, classification, targeting and actuation loop behaves under dust, heat and low light. Speed and payback figures are company-reported. No independent field trial of the G2 itself has been published, and no South African deployment is confirmed in the public record. Independent peer-reviewed laser-weeding research does exist, and it shows where the method is effective and where it is limited.

The compute stack

According to Carbon Robotics, the G2 is built around a redesigned core power and processing unit using the latest NVIDIA processors, improved lasers, perception cameras and lighting. Its software layer, Carbon AI, is powered by a large plant model trained on data collected by its LaserWeeder fleet. The operational loop is image, inference, classification, laser targeting and actuation, repeated continuously while the implement moves through the field (The Robot Report on the LaserWeeder G2; product details from Carbon Robotics).

Latency and actuation

The actuator is a bank of vision-targeted lasers that deliver thermal energy to the weed's growing point rather than a chemical or a mechanical blade. Carbon Robotics claims the G2 units are twice as fast as the earlier generation and more power-efficient. Speed matters because classification and firing must complete within the window between a plant entering the camera's field of view and the implement passing it.

Deployment envelope

Carbon Robotics reports five product configurations, with the lightest implement starting at 1,930 kg (4,250 pounds) so that smaller tractors can pull it, and a 6-metre (20 ft) version that is 25 percent lighter than the original LaserWeeder. Modular design supports widths from 6.6 to 60 feet. The company puts the payback period at 1 to 3 years on equipment with a 7 to 10 year lifetime. These are vendor-reported specifications, a reason to investigate rather than verified local performance.

What the peer-reviewed record shows

A Cornell and Rutgers study published in Pest Management Science in 2025 tested laser weeding on processing vegetables on the US East Coast (the study, DOI 10.1002/ps.8912). It found laser weeding at least as effective as standard herbicides on erect annual weeds, but weaker on purslane and annual grasses. Control depended on catching weeds while they were small and on repeated passes at 1 to 2 mph. Against untreated plots, laser treatment cut weed cover by 45 percent or more, weed density by 66 percent or more and weed biomass by 97 percent or more, with crop stunting of no more than 1 percent. The study did not test the commercial G2 directly, so its results describe the method's conditions of effectiveness, not the G2's field performance.

Why it matters for South African operators

The G2 treats weed control as a perception-and-compute problem rather than only an agronomy-and-chemistry problem. In that framing, a missed classification, a lighting change or a camera fault becomes an operational failure in the same way a blocked nozzle would be on a sprayer. The peer-reviewed record shows the method rewards early, repeated passes on small weeds, which is an operational discipline a farm must be able to sustain across the season. For a South African adoption decision, two conditions would change the trade-off: local weeding-labour cost and availability, and whether the local weed spectrum is dominated by the erect annual weeds where the method is strongest or by grasses and purslane where it is weaker. No local data yet shows how the G2's perception loop holds up under South African conditions, so those questions stay open.