01 / BRIEF
A radio-controlled car you drive over the internet, seeing through the camera on its nose. Today from a browser, next from a headset. The version that keeps me interested is the one at the end: a track with several of them, and people racing each other from their own houses.
Nobody needs this to exist. That is precisely the point: telepresence is a hard problem, and this version of it fits on a desk and fails cheaply. Everything interesting about controlling a machine you are not next to shows up here at a scale where a mistake costs a bumper.
02 / HOW IT WORKS
Two brains, split by deadline. The car carries a microcontroller and a single board computer, and they are separated on purpose. The microcontroller is the muscles: motors, steering, failsafe, a loop measured in milliseconds that must never wait for anything. The single board computer is the eyes and the network: it holds the camera, encodes the video and carries the connection. They talk to each other over a serial line, which is the cheapest, most boring link available and therefore the one that does not fail.
The line in the middle. Everything that crosses the internet is treated as unreliable by default. Command goes one way, video comes back the other, and neither of them is allowed to hold a motor hostage. If the link drops, the car does not freeze mid-corner waiting for a packet: the reflex is on its side of the line and takes over.
The return path is the enemy. Video that leaves the camera, gets encoded, crosses a network, gets decoded and lands in a headset spends milliseconds at every step. On a monitor that is lag and you adapt. In a headset your inner ear does the arithmetic, disagrees, and you feel it in your stomach within a minute. So the work is a budget: measure every stage, write the number down, and make anything that wants to add to it justify the cost.
A real place. Cold rubber grips differently than warm. A connector works until the third lap. Indoors, full throttle is useless, so the throttle ramps instead of jumping. The physical world keeps producing failures a simulator would never have shown me, which is the entire reason to build this in atoms.
03 / DECISIONS
The reflex lives on the vehicle. Anything with a deadline runs where the network cannot starve it. This is the same rule as the rest of the workshop: put the thing that must not fail on the side of the boundary you control.
A real car, not a simulation. Simulated latency teaches you nothing about the loose connector or the brownout when the motor draws hard.
Milestone one is the boring one. Throttle and steering first, from a keyboard, with the gain turned down to a quarter for the first test. Every project that starts with the headset ends with a headset and no car.
Build the useless version first. A toy has no stakeholders, no deadline and no scope creep, which makes it the most honest place to learn a hard problem.
04 / STATUS
Milestone one closed in June 2026: throttle and steering working, untethered. Camera and headset next.