Your robot

The whole course runs on one arm: the SO-101, Hugging Face's low-cost 6-DoF arm. This page is the station it runs on — the two arms, the two cameras, and the ports the kit talks to — so you know what's on the table before you drive.

Two arms: leader and follower

You work with a pair. You move the leader by hand; the follower mirrors it in real time. Recording captures the follower's motion — its joint positions as the observation, the leader's commanded goals as the action — so a demonstration is you moving the leader while the follower copies. Both connect over USB and show up as /dev/cu.usbmodem<USB_ID>; you never type those ports by hand, because calibration writes each arm's id to calibrations/<USB_ID>.json and later runs read it back — including which arm is the leader.

Two cameras: top and side

The SO-101 base model was trained on two third-person views, so that's what your rig needs: one camera looking down at the workspace, one looking across it. The kit exports the first detected camera (cam0) as top and the second (cam1) as side, in that order — check the live stream to confirm which physical camera is which before you record, and keep the placement fixed for the whole dataset.¹ Avoid a wrist-mounted camera: the base checkpoint never saw one, so it can't use one.

Recording slot — poster pending

The ports the kit uses

One long-lived local server owns your data for the whole session, and it listens on a handful of ports. You don't manage them — but if a port is busy, this is what's on it:

9876 — the live stream the kit's local site connects to. 51234 — the catalog of every recording you make. 8000 — the control API behind the record buttons. 3000 — the kit's local course site itself (pixi run learn). The server attaches to the arms only on demand, so it never blocks calibration or teleop from grabbing the serial ports.

That's the station. Head to Quick Start to install the kit, then start driving.

→ The SO-101 kit on GitHub — the README lists every station command.