6 · Train and run inference
Previous: 05 Record a datasetNext: 07 Troubleshooting
Training turns the dataset into a policy — a model mapping camera images to joint positions. Inference runs that policy on the follower arm.
Train
Open the Training tab and choose:
| Choice | Notes |
|---|---|
| Dataset | A local dataset, or one from the Hub |
| Policy type | ACT is the usual first choice for a single arm. Diffusion Policy, SmolVLA and the π₀ family are also available |
| Where to run | This machine, or a Hugging Face cloud job |
| Training length | Number of steps. Watch the loss rather than assuming more is better |
The tab streams the training log and shows checkpoints as they are written.
Local or cloud
| Local | Hugging Face cloud job | |
|---|---|---|
| Needs | A GPU on this machine | A Hugging Face account with access |
| Speed | Depends on your GPU | Fixed machine, chosen per job |
| Best for | Small datasets, experimenting | Longer runs |
The cloud option is why a laptop is enough for the rest of the workflow: record locally, train remotely.
Policy extras
Some policy types need dependencies LeLab does not install by default. Selecting one prompts you to install them — accept and wait for the background install to finish before starting the job. The same mechanism covers optional training extras.
Checkpoints
Training writes checkpoints periodically, not only at the end, so an interrupted run still leaves a usable model. Each checkpoint name carries its step number. Try an earlier one if the loss started climbing.
Inference
Open the Inference tab:
- Confirm both arms are connected and calibrated.
- Select the policy — a local checkpoint, or a model from the Hub.
- Confirm the cameras and their rotation.
- Start inference, and clear the workspace.
Camera rotation and binding
Use the same rotation as during recording. This is the most common reason a well-trained policy does nothing useful on the real robot, and nothing warns you: 0° and 180° produce identical image dimensions.
The same applies to the camera binding. If the camera called wrist during recording is now a different physical camera, the policy sees the wrong image.
If all of that is right and the arm still fails, look at the dataset again — see page 5.
Intervening
Stop inference before reaching for the arm. A correction you make by hand is not training signal unless you record a new episode and retrain.
Replay
Replay re-runs a recorded episode on the follower arm. It plays back the recorded joint positions, so it looks correct even if the policy is bad — it checks the hardware and the dataset, not the policy. Use inference to see what the policy does.
Next: 07 Troubleshooting