Two things carry through every step:
- The task text, like
pick up the cube. You set it when you record. The policy learns to do that task. - The Training ID. You copy it from the dashboard after training. The inference script uses it to download your policy.
Before you start
- Two SO-101 arms, calibrated. See SO-101.
- An auth token and a room id. Use the same ones on both machines. Cloud recording is billed to the token’s account.
- Python 3.12 or newer, and the examples repo:
Related
Cloud recording
Step 1: record demos.
Datasets
Step 2: turn a recording into a dataset.
Fine-tuning
Step 3: train a policy.
Running a policy
Step 4: let the policy drive the arm.