Train a skill

Tutorial 4 ran someone else's checkpoint. This one makes a checkpoint, first from a dataset you did not record, then from demonstrations you did.

The training command is local: solarian train, on the machine where you installed the trainer. Solarian Base can run the same config on a hosted machine. This tutorial does not. One path is enough, and the config file is the same either way.

You need Tutorial 4 through the pretrained pick, because the last step is pointing that same skill at your new file. Collecting demonstrations needs the Lumi hardware. Training on the example dataset does not.

Chapters

1. Datasets — episodes, what a useful demonstration looks like.
2. An example dataset — train before you own a single episode.
3. Run your checkpoint — the pick skill, aimed at the file you just wrote.
4. Lumi — the hand-held recorder, what it is and how to build it.
5. Collect — episodes, packed, then turned into a training set.
6. Train on your data — drop the short episodes, train, run it on the robot.

What you are producing

A training run writes models/<job_name>/checkpoints/latest.ckpt. The skill YAML's policy_checkpoint points at that file. Nothing else about the controller changes. Same key as Tutorial 4. Different weights.