SO-101 + LeRobot + Odyssey-3 Robot Arm Setup Workflow

Step 1: Get four SO-101 arms, two you drive by hand and two that copy them, $278 a kit plus $30 in printed parts. Step 2: Mount one camera overhead and one on each wrist, then bolt them down and never move them again. The model reads pixels, so a shifted camera invalidates every episode you already recorded. Step 3: Install LeRobot from Hugging Face; these arms are natively supported for teleop and imitation learning. Step 4: Before recording anything, open Odyssey-3 and prompt your exact task. It's free right now and you'll find out in 10 minutes whether its physics holds for your objects. Step 5: Move the leader arms yourself and record, keeping the frame rate fixed and the lighting constant across every session. Step 6: Pick one task and go deep. Their arm results used tens of hours, and ten hours of pouring beats one hour each of ten different tasks. Step 7: Keep the failed runs. A policy that only ever saw clean episodes has no idea how to recover from a bad grip. Step 8: Train a plain policy in LeRobot first. That's your baseline, and without it you can't tell whether Odyssey actually helped. Step 9: When the API opens, the world model stays frozen and you train a small action decoder on top, mapping your recorded observations to your arm's joint commands. Step 10: Expect a gap on first contact. Their driving policies trained purely in sim reached about 77% of what real footage produced. How the three fit together: - SO-101 produces the actions and executes them. - LeRobot records the observation and action pairs, then trains on them. - Odyssey-3 is the physics you attach to instead of learning it from scratch. Your episodes are the only part that has to be yours; everything else is already trained.