Last month I bolted a Xiaomi Wowstick screwdriver to a robot arm that costs less than dinner for two. The arm drove the bit into a screw, missed by 3 mm, tried again, and sank it flush. The whole thing sat on my kitchen table in Vienna. No safety cage. No six-figure budget. Just a 58-gram screwdriver, a 3D-printed clamp, and an SO-100 arm running LeRobot.

That's when the idea clicked.

The hardware

Four SO-100 arms. Two leader-follower pairs — you grab the leader, the follower copies your movements. I also have a Dremel rotary tool and a cheap airbrush spray gun from Amazon. The screwdriver already mounts (barely). The other two need adapters I haven't printed yet.

Why four arms? Because two pairs let you build a chain. Follower A picks a part off a tray. Follower B drills a hole in it. That's a conveyor. A terrible, wobbly, slow conveyor — but a conveyor. And when a 16-year-old sees two robots passing a block between them and doing something to it, something changes in their head. I've watched it happen.

What I want to do with this

I'm planning a course. Eight sessions, 90 minutes each, for kids 14 and up. The arc goes roughly like this: first session they assemble the arm from a kit. By session four they're recording demonstrations and training an imitation learning policy. By session seven they've got both follower arms running in sequence — one feeds, one drills or paints. Session eight is demo day. They show their mini-factory to whoever wants to watch.

I don't have the lesson plans written. I don't have safety protocols drafted. I haven't recorded a single trajectory in LeRobot format. This repo is the napkin sketch, not the blueprint. I want to run through the whole thing myself at least once before I write it up for anyone else.

Why this matters here

Austria has a manufacturing sector that punches above its weight — Voestalpine, Andritz, hundreds of Mittelstand companies — and a growing shortage of automation engineers. The technical universities teach robotics, but on hardware that costs more than a used car, in labs you can't take home. I want something a Gymnasium student can set up in their bedroom for the price of a video game console.

And there's a second angle. LeRobot is a HuggingFace project. Every trajectory you record gets stored as a dataset on the Hub. So these kids wouldn't just learn robotics — they'd learn the HF ecosystem too. Record a demo, push it, train a policy, pull it back. The whole loop.

I might be overestimating what a 15-year-old can absorb in 90 minutes. Wouldn't be the first time. But I'd rather aim high and adjust than build something too simple to be interesting.

What's not done (basically everything)

Tool mount adapters — not designed. Safety docs — not written. Curriculum — in my head, not on paper. Recorded trajectories — zero. Gradio demo — maybe eventually.

If you teach robotics, work with SO-100, or just think this is a good idea (or a bad one, honestly), open a discussion. I could use the feedback.

Downloads last month: 12