In twelve weeks there's a machine on your floor that you assembled, wired and programmed. You tell it to go to the kitchen and it goes — around the chair, past the dog, without you touching it. You know why every wire is where it is, and when it stops working you know where to look. Then it does a job you chose for it.
Not three disposable projects. One machine that gets more capable every month — which means you're never starting over, you're always extending hardware you already understand. Here's what you can do at the end of each month that you couldn't at the start of it.
You've soldered it together, written the firmware, and told it to drive a one-metre square — and it comes back to where it started. You know what an encoder tick is because you counted them, and you can make a motor do exactly the speed you asked for.
You wrote the filter that fuses wheel odometry with raw inertial data — not configured someone else's, wrote it. Your robot now holds a map of your actual floor, knows its place in it, and drives to any point you name.
It sees things and decides what to do about them, running a vision model on its own accelerator. You pick the job — fetch, patrol, follow, inspect — and you build in the e-stop, the watchdog and the battery monitoring that let you walk away from it.
One robot, but three separate documented results — so you're not a candidate for one narrow role, you're a credible applicant to three different kinds of robotics team. Each one is a repo, a write-up and a video, and every claim in it is backed by a number you measured yourself.
Firmware, encoder handling, closed-loop velocity, wheel odometry — with the measured square and the plots to back it. Hand this to an embedded or controls team and they can see you've made hardware obey you.
Your own estimator, lidar SLAM, and Nav2 tuned for a real floor with real furniture — reported as a success rate over repeated trials, not a single lucky run.
A quantized vision model on an accelerator inside a behavior tree, with a latency budget, thermal numbers and a thirty-minute unattended run. Very few applicants have measured any of that.
Why this matters more than the robot itself. Interviewers can't tell a hard project from an easy one by looking at a video. What they can assess is whether you measured anything, whether you know why it failed the other two times out of ten, and whether you can defend a number. That's what these three are built to give you, and it's why every stage ends at a bar you have to actually hit.
Everyone in the cohort builds the same machine — a differential-drive robot with a lidar, a camera, an IMU and an AI accelerator on board. It's the same shape as the machines that run in warehouses, only smaller and on your floor.
It costs about $450 on top of the program, and that's real money, so know what you're getting for it: a working robot that stays on your desk afterwards, plus a soldering iron and a multimeter you'll use for the rest of your life. Nothing is rented and nothing is returned.
You will not be hunting for parts. When you enroll I send exact part numbers, links, and a hard order-by date, so nothing is stuck in a shipping queue when week one starts. If something you ordered arrives dead — and occasionally it does — that's a message to me, not a lost weekend.