quiet machines

most machines are loud. not in decibels. in the attention they demand. a robot that needs an integrator to install it, an engineer to retune it after it is moved two metres, and a specialist to explain why it stopped at three in the morning is a loud machine, however quietly its motors run.

i think this is the real cost of hardware today, and i think it is mostly self-inflicted. this essay is my thesis on how to build the other kind.

ithe body is not the hard part anymore

a $250 arm on a desk now does what a lab arm did a decade ago. actuators, encoders, cameras and compute have all become commodities. you can order the parts for a capable machine on a sunday and have them by friday.

what has not become a commodity is the part that makes a specific machine do a specific job. every machine still ships with its own brain, and every brain is still written more or less from scratch: a custom controller, a custom calibration routine, a custom stack of assumptions about this gripper, this camera, this table. the hardware got cheap. the glue did not.

iibuild the body boring

so the first half of the thesis is about restraint. build the body boring.

standard actuators. standard interfaces. parts you can replace without redesigning anything around them. no mechanism so clever that only one controller, written by one person, can exploit it.

cleverness in the mechanism is a debt the software pays forever.

a beautiful custom linkage saves a few grams and a few percent of cycle time, and in exchange it makes the machine a one-off. nothing learned on it transfers anywhere. a boring body is an invitation: anything that can drive one of these can drive all of them.

iiiput the intelligence where it can travel

the second half is where the effort should go instead. a model that controls one machine is a product. a model that can control any machine is infrastructure.

that is what i mean by physical-use models: one foundation model that learns the structure of acting in the world, not the quirks of a single body, and so can sit behind a $250 lerobot arm and a $50M factory line alike. the cheap arm matters more than it looks. it is where data is cheap, mistakes are cheap, and a thousand people can try a thousand things. what is learned there should carry upward. the factory is where it pays off.

this only works if the bodies are boring. portable intelligence needs predictable hardware underneath it. the two halves of the thesis are one idea seen from both sides.

ivwhat quiet means

a quiet machine is one you describe a job to. it does the job, and it tells you plainly when it cannot. no teach pendant. no week of integration. no retuning because the light changed.

the measure is simple: how little you think about it. good hardware disappears into the work the way a good tool disappears into the hand.

vwhat this costs

none of this is free. generality is slower to build than a hand-tuned cell. early demos look worse than a system engineered for one task. you give up the last few percent of speed that a bespoke controller could squeeze out.

my bet is that those few percent matter far less than the months of integration they cost, and that the gap closes from below: the general system keeps improving with every machine it touches, while the bespoke one is as good on its last day as on its first.

build the body boring. make the mind portable. let the machine be quiet.

---

poznań, september 2026