Start with the task, not the farm
Agriculture contains many different physical jobs: scouting crops, identifying weeds, applying a treatment, milking animals, thinning fruit, and harvesting. A robot is easier to justify when it removes a specific access, timing, or repetition problem. A 2026 USDA specialty-crop program report describes automation across sensing, packing-house inspection, assisted harvesting, weeding, spraying, and vineyard treatment; it does not imply that one platform can do all of them.
The crop and production system shape the machine. The methods for a USDA-funded nursery project call for measuring the target crop, lighting, required force, and approach before choosing cameras, range sensors, controls, manipulators, and end effectors. That is a useful boundary for a founder: “agriculture” is too broad a product definition; one crop, operation, growing system, and operating window is a testable starting point.
What has been demonstrated
Automatic milking is a deployed category rather than a lab idea. A January 2026 USDA Economic Research Service report uses US farm survey data to study robotic milking alongside breeding and herd-management technologies, and says adoption of precision dairy tools has risen since 2000. Its profitability result is an econometric estimate across surveyed farms, not a guarantee for a particular dairy or a general benchmark for agricultural robots.
Crop robots have also moved into field trials, but trial scope matters. University of Georgia researchers reported a 2024 autonomous laser-weeding experiment in one cotton field. Their best reported single-pass result used weed tracking, while the paper still calls for better effectiveness, speed, reliability, crop protection, ecological evaluation, and laser-safety rules. This demonstrates an integrated prototype under stated conditions, not season-long commercial performance.
Where the hard work remains
Outdoor perception changes with illumination, weather, uneven ground, plant growth, and occlusion. Manipulation adds crop damage, reach, cycle-time, and recovery questions. A 2024 kiwifruit-harvesting paper is a useful caution: later picking tests used a laboratory trellis because the harvest season constrained the experiment. Published component results can therefore be meaningful without proving dependable work in a living crop over a full season.
What building one involves
A credible build plan ties the whole loop together: detect the right biological target, localize the vehicle and target, move through the site, apply the tool without harming the crop, record what happened, and let an operator stop or recover the system. Field data must cover the actual varieties, growth stages, light, terrain, and weather the product claims to handle. Maintenance, cleaning, transport, charging, and seasonal access belong in the first pilot plan because a successful lab action alone does not establish a workable farm operation.
Where Vesk fits
Simulation evidence for a bounded ROS 2 task.
Vesk can help generate and test ROS 2 Jazzy code for a defined agricultural workflow in controlled simulation. That evidence is simulation-only: it does not assure field reliability, safe physical motion, sensor performance, crop protection, or hardware readiness.