The data physical AI is missing.
Everyone can train models. Almost nobody has real-world data from machines working in unstructured outdoor terrain. We spent ten years collecting it, and it grows every day.
multi-sensor, multi-season field data
expert annotations
robots adding new data daily
corn, soybean, sorghum, cotton, rice, table grapes, oil palm
Action-conditioned, or it does not count.
Most embodied AI data comes from tabletop manipulation, indoor teleoperation, or driving on mapped roads. Internet video shows what the world looks like, not how it responds.
Action-conditioned data pairs every observation with the command executed and what happened next: the slip, the sink, the correction. That is what a world model learns dynamics from.
Every terabyte in the corpus came from a robot doing real work in a real field.
Simulation is excellent at geometry and terrible at dirt.
Deformable terrain, wet vegetation, heat shimmer, canopy light: the conditions that break outdoor autonomy are the ones simulation approximates worst.
Synthetic data is abundant. Ground truth from the physical world is not: it takes fleets, field operations, and seasons.
What ten years in the field looks like.
Environments
Under canopy, heat distortion, deformable terrain, dust, mud, and rain. GPS-denied throughout.
Sensors
Camera, depth, inertial, and platform telemetry, aligned with executed actions.
Seasons and crops
Multi-season coverage across corn, soybean, sorghum, cotton, rice, table grapes, and oil palm.
Annotations
10M+ expert annotations, built for training rather than evaluation alone.
Licensed by field of use.
Non-exclusive, scoped by field of use, for teams training world models and embodied AI.
Structures range from evaluation subsets to multi-year partnerships. Every conversation starts under NDA with a scoped sample.
Talk to us about the corpus.
Tell us what you are training and the environments you care about. We will scope a sample against it.