Day 4: back to real engineering, and the first casualty was Day 1

Another day starting in the virtual conference room instead of the workshop. I’ve been here before. Oh well — chalk it up to the process.

This time we actually got somewhere, though. The single joint we’ve been tracking with is now officially the starting line, not the whole plan. The real target for this first release is a full three-joint arm, broken into stages: map the 3D space around it, add a second joint and get both moving together, teach it to learn and replay what it’s seen, then bolt on a third joint with some real movement optimization on top. It didn’t exactly light up the path. But it strung a thin line of fireflies along the trail we’re planning to walk through the dense forest of questions still waiting for us — and that’s something we can actually work with.

Good news: we can finally get back to real engineering instead of talking about it. Bad news, arriving roughly four seconds later: the first thing engineering did was tell us Day 1 was wrong.

The motor we picked on day one — the MG996R, the one an entire joint got built around — turns out to be too weak for the shoulder it was actually meant for, and it has zero position feedback. No feedback means no way to know where a joint actually ended up, which means no way to record a movement and play it back later, which was more or less the entire point of teaching this thing to learn. So: new motor. The Feetech STS3215 — a size up, a real encoder, a sturdier communication protocol, still inside hobbyist money. Which is great, except its dimensions have nothing in common with the old motor’s, which means the case that ate all of day one’s work doesn’t fit it anymore. Everything from Day 1, crossed out on the spot. But that’s what a prototype is for. And nothing takes away the feeling of that first win — that feeling is what actually keeps us going.

Of course the new motor is also too big for the smaller joints — elbow, wrist, and whatever we end up doing about fingers, which is its own unsolved problem. Best guess so far: tendons, cable-driven from the forearm, the same way it works in an actual hand — the muscles for your fingers aren’t in your fingers. For elbow and wrist there’s a smaller sibling motor, the STS3032, on the same communication protocol — one standard for the whole arm instead of three.

Then came power, which turned out to be a repeat lesson: a battery and a bench supply are not the same purchase, and buying the wrong one first is how a wiring mistake becomes an expensive one. I’m afraid our robot is going to refuse to run on eggs and bacon — it needs “clean” energy, and clean energy that won’t fry a servo on the first mistake means a current-limited bench supply before anything else. Skip that step and get the wiring wrong, and I’d become a living illustration of the saying about the miser who pays twice — or three times, or four. The arm at full strength wants something like 15-20 amps, which no reasonably priced bench supply actually hits, but a Riden RD6012PW S800, rated for 12, was close enough to call it decided.

New day, and the shopping list is $120-150 heavier again. Also, somehow, we’re a few real steps closer to knowing what we’re actually building.