Recording and training

Record demonstrations by teleoperating the robot, then train a policy on them and run it back.

🚧 To write

• Recording a dataset. Starting/stopping an episode, what a good episode looks like, how many you need before training is worth attempting.
• Editing a dataset. Dropping bad episodes.
• Training. Training is cloud-dispatched — it does not run on your laptop, and the desktop app deliberately doesn't ship a local training runner. Cover: launching a job, what the live monitor shows, how long to expect, and reading the loss curve.
• Training history and picking a checkpoint.
• Inference. Inference runs locally — torch ships inside the desktop bundle on purpose, so the motor-command loop never depends on Wi-Fi. Cover: loading a policy, running it, stopping it safely.
• Marketplace. Publishing a policy and using someone else's.

Why training is cloud and inference is local

Worth stating, because it explains two things people ask about:

Inference is local — which is why the desktop download is ~770 MB rather than ~200 MB. A robot's motor-command loop must not depend on your Wi-Fi holding up. That size is a deliberate trade, not an oversight.

Training is cloud — it needs GPUs you don't have, and it's not latency-sensitive. Nothing about training touches the robot.