LeRobot Command Builder

1

Calibrate

Teaches LeRobot each joint's range. Do the follower and the leader separately.

What to do physically

Move every joint to its middle position, then press Enter.
Sweep each joint slowly through its full range, one at a time.
Press Enter to save.

Calibration is stored per id. If you change the id later, LeRobot will ask you to calibrate again — so pick a name and keep it.

When to redo it

You reassembled or reprinted a joint.
You swapped a motor.
Teleoperation feels offset even though the ports are right.

Teleoperate

No recording. Prove the whole chain works before you spend an hour on data.

Control loop FPS

3

Record a dataset

Arrow keys drive the session: → next, ← redo, ESC finish & upload.

Dataset name

user/

Task description (one sentence) Episodes Episode seconds Reset seconds Dataset FPS

Push to the Hub when finished Private dataset Resume an existing dataset

Then check your data — before training, always

Open dataset viewer ↗ Opens the Hub viewer for this dataset.

The golden rule

Your first 50 episodes should be a constrained version of the task: one object, roughly fixed position, fixed cameras, one operator. Train a quick ACT model, see what fails, then widen.

Data quality beats model choice

Same grasp, same approach vector, same timing every episode.
Deliberate and clean beats fast and sloppy. Speed comes last.
Diffuse, consistent lighting matters more than resolution.
Bolt down the rig and the cameras before recording anything.

What to look for in the viewer

Dropped frames or FPS dips.
Motion blur or an out-of-frame gripper.
Episodes where the task actually failed — delete them.

Train

Pick a policy, point it at a dataset, decide where the weights land.

Train on dataset

Click to pull in the dataset you just recorded, or type any LeRobot-format Hub dataset — it does not have to be your own. Policy Device Batch size Steps Output directory (local checkpoints) Push trained policy to Leave empty to keep the policy local only. Run on Hugging Face Jobs (optional) No local GPU? Put a flavor here. List them with hf jobs hardware.

Log to Weights & Biases Resume from last checkpoint

Policy notes

How long

ACT, ~50 episodes: 50k–100k steps is a reasonable first run.
Watch the loss curve, but judge on the real robot — loss is a weak proxy for success rate.
Out of VRAM? Halve the batch size before you touch anything else.

Checkpoints

Checkpoints land in outputs/train/<job>/checkpoints/. The one you want at rollout time is checkpoints/last/pretrained_model.

Rollout

Run the policy on the real robot — and measure it honestly.

Policy path or Hub repo

Click to fill in where step 4 saved your policy, or paste a local checkpoint path or Hub repo id. Strategy Inference backend Duration (s) 0 = run until you stop it. Rollout FPS Record rollouts into

user/

Interactive mode (/start, /subtask, /stop from stdin)

Strategy

Inference

Evaluate like you mean it

Run ~10 attempts under the same conditions as training and count successes. That number, not the loss curve, is your result.
Compare against a teleoperated baseline — if you can't do it by hand reliably, the policy won't.
Failures cluster. Watch where: wrong approach angle, grasp slips, object outside the training distribution. Each one points at different data to collect next.

Want those attempts saved? Switch the strategy to episodic or sentry — they record while the policy runs. Feed failures back with dagger: you take over when it goes wrong, and only your corrections get recorded.