Anatomy of Balance.
The same printed humanoid, taken apart three ways — the body that stands, the loop that catches it, and the brain that decides where to step. Where the drone paid for every second in the air, this one stands for almost nothing; where the drone leaned to move, this one steps to a spot it can name before its foot lands. That is agency in miniature: standing for almost nothing is the viability floor, and naming the foothold before the foot lands is the certificate — the same energy the body descends, proving the step before it commits. Every number on the board is computed by the Institute's open balance core, so the page cannot lie to you. The same pedagogy as Anatomy of Flight, on a machine that could not be more different.
Guess before you look.
Each opens with a question almost everyone gets wrong, a humanoid you can turn over in 3D, and the one move that makes it obvious. Open them in order, or jump to the layer you want.
What holds it up?
“A drone burned 200 W just to hover. What does standing cost this robot?”
Take the humanoid apart: heavy low pelvis, jointed legs, wide flat feet. The surprising number is the one the body spends to stand — and it is nearly nothing. Standing does no mechanical work; the legs are a structure carrying the weight, not a motor lifting it.
Open the demonstration →How does it catch itself?
“Shove it hard from behind. What does it do first to not fall?”
The loop that keeps it up — sense, predict, step. It has one clean rule you use every day: to stop a fall, put your foot on the capture point, ξ = x + v/ω. The controller on the board steps there live; a gentle nudge it just holds with the ankle.
Open the demonstration →What decides where to step?
“A neural network keeps this robot standing. How big is its brain?”
Something runs the ankle a hundred times a second. Here it is a network — 337 numbers — balancing the real body live. It learned the ankle by watching the rule, with the step left as a reflex underneath: shove it and it catches itself, at a cost you can count in nanojoules.
Open the demonstration →One format, a second machine.
This is the same engine that runs Anatomy of Flight — the same withhold-guess-prove beats, the same editable source beside the board, the same rule that nothing is animated by hand. Only the subject changed: from a thing that stays up by pushing air down to a thing that stays up by not falling over. If the format travels, it is a format.
Guess first
A real question, refused an answer, until you commit. The drone lesson asked how long it flies; this one asks what standing costs — and the honest answer, about nothing, is the whole point.
It cannot lie
Every claim is checked live by the open balance core: the capture point that stops a fall, the ankle-only limit of half a metre per second, the 337-parameter policy recovering 24 of 24 shoves. Verified, not asserted.
Hardware to brain
One humanoid carries the whole stack: the mechanics of the standing body, the capture-point loop that catches it, and the learned policy that runs the ankle — the three layers of an embodied system, again, in one place.
Start with the body →Take the course: Embodied AI Design →← See Anatomy of Flight