✳ Field Notes · 2026
Notes written before the robots exist.
Three essays on why we chose this direction. No results to report yet — just the thinking we're willing to be held to.
Note 01
Why hands are the hard part
MOYU · 2026 · 4 min
A robot can walk across a stage, do a backflip, and still be unable to hand you a cup of tea without crushing it. That asymmetry is not an accident. Locomotion happens between a machine and the ground — one surface, one contract, endlessly repeatable. Manipulation happens between a machine and everything else.
Every object a hand meets is a new physics problem. An egg and a wrench can sit side by side on the same table, demand the same reach and the same closing of fingers — and tolerate completely different forces. The information that separates them is not in the geometry. A camera can tell you where the egg is. It cannot tell you how hard you may hold it. That knowledge lives in experience, in touch, in consequences.
Walking is a conversation with the floor. Grasping is a conversation with the world.
This is why we believe hands are where general-purpose robotics will be won or lost. The industry has made spectacular progress on bodies that move. The bottleneck — the thing that keeps robots in demo videos instead of workshops — is the last thirty centimeters: the judgment to know how hard to hold, and the sense to know what comes next.
That judgment is what we've set out to build. Not a hand — the brain that runs one. Vision to see the egg, language to be told "gently", action to close the fingers, all in a single loop that learns from every attempt. We haven't built it yet. This page exists so you can watch whether we do.
✳ MOYU
Note 02
Data quality for manipulation
MOYU · 2026 · 4 min
The lesson of modern AI is brutally simple: models are downstream of data. For language models the internet existed; for manipulation, nothing comparable does. Every team in this field is, in effect, deciding what its model will believe about the physical world — one demonstration at a time.
That makes data quality a moral choice disguised as an engineering one. A thousand sloppy demonstrations teach a model that sloppiness is the task. A teleoperator who drags an object across a table because it was faster that day has just taught the model that dragging is grasping. The model will not know the difference. It was never told there was one.
A model trained on careless demonstrations doesn't learn the task. It learns the carelessness.
So our position — stated here before we collect our first demonstration, so it can be checked against what we actually do — is that we will go slow on purpose. Fewer demonstrations, recorded properly: clean task definitions, deliberate motions, labels that say not just what happened but whether it should have happened that way.
We'd rather have a hundred demonstrations a reviewer can defend than ten thousand nobody has watched. The bitter lesson says scale wins; our bet is that for manipulation, scale of verified experience wins. The difference between those two sentences is where we intend to live.
✳ MOYU
Note 03
Evaluating grasps honestly
MOYU · 2026 · 3 min
Robotics has a demo problem. A grasp that works once, on camera, with the lighting just right and the object placed by the person filming, is not a capability — it's a coincidence with good production values. Everyone in the field knows this. The incentives to pretend otherwise are enormous.
The honest question is never "can it pick this up?" It's "out of a hundred tries, on objects it has never seen, placed by someone who doesn't care whether the demo succeeds — how many times?" That number is usually embarrassing. Which is exactly why it's the only number worth publishing.
A benchmark you can't fail is marketing. We intend to publish the failures.
When our bench exists, every capability we claim will come with that number: success rate, over how many trials, under what variation, with failure videos alongside the successes. If a grasp works 73 times out of 100, the page will say 73 — not "demonstrates robust grasping".
This note is a commitment device. It costs nothing to write today, when there is nothing to evaluate. The day it costs something is the day it matters. Hold us to it.
✳ MOYU