Anatomy of Driving.
The fourth machine, and the cheapest way to move — a rover, taken apart three ways: the wheels that carry it kilometres on a sip of charge, the one rule that forbids it from going straight sideways, and the brain that turns a parking spot into steering and throttle. This is physics the drone, the humanoid and the hand never met: the nonholonomic constraint, the reason a wheel is free to roll and forbidden to slide. Every number on the board is computed by the Institute's open drive core, so the page cannot lie to you.
Guess before you look.
Each opens with a question almost everyone gets wrong, a rover 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.
How far does it roll?
“The drone's battery hovered it 11 minutes going nowhere. On wheels, how far does the same charge take it?”
Take the rover apart: wheels, battery, a sensor mast. The surprising number is the range, and it is enormous — a rolling wheel barely fights the ground, so distance is energy over rolling resistance, and the same charge that held the drone in the air goes tens of kilometres. The wheel is the cheap way to move.
Open the demonstration →Why can't it go sideways?
“Tell it to move one metre directly to its left. What does it do?”
The rule you have felt every time you parallel-parked: a wheel rolls along its heading and never sideways, so ẋ·sinθ − ẏ·cosθ is always zero. There is no input that points left. To reach a pose beside it, the rover drives a curve — out, around, and back in — the shortest path it is allowed.
Open the demonstration →What decides how to park?
“A neural network parks this rover. How big is its brain?”
Something turns a target pose into steering and throttle a hundred times a second. Here it is a network — 746 numbers — driving the rover into the spot live. It learned the parking curve by watching the controller, with the nonholonomic constraint left exact underneath: it must find the curve, and it does, spot after spot.
Open the demonstration →One format, a fourth machine.
The same engine that runs Flight, Balance and Grasp — the same withhold-guess-prove beats, the same editable source, the same rule that nothing is animated by hand. Only the physics changed: from staying up, catching a fall, and holding on, to the constraint that lets a wheel roll forever forward but never a metre to the side. Four machines now, and each is useful for a reason none of the others share.
Guess first
A real question, refused an answer, until you commit. This one asks how far a battery drives after it could only hover — and the answer, tens of kilometres, is the surprise that opens the door.
It cannot lie
Every claim is checked live by the open drive core: the range from rolling resistance, the sideways velocity that stays exactly zero, the 746-parameter policy parking the rover on 23 of 24 spots. Verified, not asserted.
Hardware to brain
One rover carries the whole stack: the mechanics of the rolling body, the nonholonomic rule that shapes every path, and the learned policy that parks it — the three layers of an embodied system, a fourth time.
Start with the body →Take the course: Rust for Physical AI →← All six machines