Physical AI

Intelligence that learns the work, then improves it.

A picking robot is only as good as its judgement: which fruit is ready, how to reach it, how much force is too much. We treat that judgement as something to be learned from people and sharpened by experience — not hand-coded crop by crop. This is physical AI: software that acts in the real world, and gets better every time it does.

From demonstration to autonomy. A skilled picker shows the robot how the job is done. The robot turns those demonstrations into its own behaviour, tries the task under supervision, and folds every correction back into what it knows. Over a season, it moves from copying to competence.

The principle that ties it together is common sense in the field — handling the unexpected fruit, the awkward stem, the wet morning — with the same adaptable judgement a person brings, rather than a brittle script.

Three words for what we build

  • Autonomous — it takes responsibility for the task, not just a single motion
  • Self-learning — it acquires skill from people and from doing the work
  • Self-improving — it gets measurably better with every shift and every robot
How the intelligence works

Five ideas behind the mind.

A high-level look at the thinking. The specifics of our models, data and methods stay in-house — but the shape of the approach is something we are glad to share.

Learning from people

Experienced pickers demonstrate the task — including with a motion-capture glove — and the robot learns manipulation the way an apprentice would: by watching the people who know the crop best.

Imitation learning

Common-sense reasoning

Real rows are messy. The robot is built to reason about what it is doing — to handle the bruised fruit, the tangled stem, the unexpected angle — instead of failing the moment conditions leave the script.

Adaptable judgement

Self-improving

Every pick is feedback. When the robot is unsure it asks for help, and that correction becomes training. The machine you deploy in week one is not the machine you have by harvest's end.

Improves with use

Fleet learning

Robots do not learn alone. Lessons earned by one machine can lift the whole fleet, so the platform's skill compounds with every unit in the field — a head start that grows as we scale.

Swarm intelligence

Seeing beyond sight

On-board 3D and advanced imaging let the robot judge ripeness and plant health in ways the human eye cannot — turning perception into earlier, better decisions about what to pick and when.

Edge perception

Safe by design

Autonomy and trust go together. Hardwired safety, confidence-gated action and a clear record of every decision mean the robot can work near people and prove what it did.

Trustworthy autonomy
The learning loop

Demonstrate. Try. Correct. Improve.

The same loop runs whether it is the first day on a new crop or the thousandth row of the season.

01

Demonstrate

People who know the crop show the robot the job — the approach, the grip, the choice of which fruit is ready. Good demonstrations are the seed of good behaviour.

02

Try, supervised

The robot attempts the task with a person watching. It acts when confident and pauses when it is not, so nothing is learned the wrong way.

03

Correct

Every intervention is captured as feedback. The robot does not just get told it was wrong — it learns the right thing to do next time.

04

Improve & share

Skill earned in the field updates the model, and lessons can lift the wider fleet. The platform compounds what it knows.

Questions

What people ask about the AI.

No. The whole point of learning from demonstration is that the people who already know the crop can teach the robot. A farm technician starts and stops shifts; the robot handles the picking and the learning.
It is built to reason rather than follow a fixed script, and it improves with exposure. New light, new weather, a slightly different plant — these are things it adapts to, and each one makes it more capable.
It is honest about uncertainty. When confidence is low it pauses and asks for help instead of guessing, and that moment becomes training data — so uncertainty shrinks over time.
Our framing is acceleration, not replacement. The robot does the relentless, hard-to-staff picking work; people teach it, supervise it, and move up to higher-value tasks. The crop that would have rotted gets harvested.

Build the mind of agricultural machinery.

Robotics, perception, imitation learning, fleet software. If you want real machines learning in real fields, we have a seat for you.

See open roles →