Quaternion Studios — the field
No thesis. No roadmap. A small group working on problems we find worth the attention — and very little else.
What surrounds you is not a services page. It is six things we believe about building machines that act in the world, placed on a ring so that none of them gets to be the headline. You have to walk to each one. That is the point.
Six positions, all 2400 units out. Scroll to walk the ring, drag to leave it ↓
Position 01 / 06
A model that only ever emits text is graded by people who already share its assumptions. Put the same weights on an arm and the grader becomes physics, which has never once been persuaded by a good explanation.
That is why almost everything we find interesting has an actuator at the end of it. Not because robots are fashionable, but because the feedback is honest and arrives on its own schedule.
What it costs — It is slower. You cannot fix a gripper by rerunning the cell.
Position 02 / 06
A policy that is right two hundred milliseconds late is not slightly worse. It is wrong — the world has moved and it is now acting on a photograph. Accuracy measured off the clock is a story about a different robot.
So the control loop is not a deployment detail you get to after the research. It is the specification the research has to fit inside. Three billion parameters inside twenty milliseconds, or it does not exist.
What it costs — We turn down approaches that are more accurate on paper.
Position 03 / 06
Almost nothing worth doing survives being written down as a reward function. The interesting part of a task is the part nobody can articulate: the hesitation before a fragile grasp, the regrip, the decision to stop and look again.
Demonstrations carry that. We would rather spend a week collecting two hundred good ones than a month arguing about a spec that will be wrong in the first unfamiliar kitchen.
What it costs — You inherit the demonstrator's habits, including the bad ones.
Position 04 / 06
A benchmark is a way of agreeing not to look. There is the number in simulation, then the number on the bench, then the number after a wet floor, a new operator, and a Tuesday — and only the last one is about your system.
We build the evaluation before the model, because the failure you cannot measure is precisely the failure you will ship. Trajectory-level, adversarial, and run often enough to be boring.
What it costs — The first honest number is almost always worse than the demo.
Position 05 / 06
A system only one person understands is a liability with a good demo attached. Cleverness that cannot be read is a bill someone pays later, usually at three in the morning, usually not the person who wrote it.
So: traces, budgets, named failure modes, and code shaped like the problem. We would rather your engineers own the thing in six months than that you keep paying us to be the only ones who can read it.
What it costs — Boring architecture. We lose arguments about elegance on purpose.
We argue all six of these at length in the journalPosition 06 / 06
Every object drifting across this plane is a raymarched fractal, evaluated per pixel, in real time. This one is yours. Drag to rotate it, push the fold and the detail past where they should go, cycle the palette until it looks wrong.
We build things like this constantly and for no particular reason, which is the credo at the origin taken literally. It is also the same renderer we ship — the toy and the product are not different code.
Palette
What it costs — Time we could have spent on something with a customer attached.