# Data Preparation

Raw recordings may be sparse, contain failed episodes, or lack visual variety. Here you clean,
label, convert, and inspect them as a LeRobot dataset.

## Learning Objectives

By the end of this lesson, you’ll be able to:

- **Explain** the four stages that turn raw recordings into a training-ready dataset.
- **Run** the preparation pipeline from a prompt and **review** the result in the visualizer.

## Run It With a Prompt

<figure style="text-align:center; margin:1.5em 0;">
  <video controls muted loop playsinline preload="metadata" style="width:100%; border-radius:6px;">
    <source src="videos/data_conversion_scissors_and_review.mp4" type="video/mp4">
    Your browser does not support embedded video.
  </video>
  <figcaption>
    Preparing the <code>scissor_pick_and_place</code> dataset: the recorded episodes are converted
    to a LeRobot dataset, then opened in the LeRobot visualizer. Camera videos appear on the
    left and synchronized joint-state and action timelines on the right, so you can confirm the
    demonstrations look right before training.
  </figcaption>
</figure>

### Expand and Visualize the Dataset

```text
Mimic 3 more episodes and visualize my dataset.
```

### Annotate Episodes for Success

```text
Run Annotation on all recorded episodes and summarize.
```

## What Happens Under the Hood?

Preparation is a short pipeline. The shipped skills cover mimic, annotate, replay, convert, and visualize; the middle stages are optional and repeatable, and conversion is the hand-off to training.

![Pipeline from raw HDF5 through expand (mimic), label and filter (VLM), optional augment (Cosmos), convert (LeRobot), and visualize.](agentic-workflows/images/data-prep-pipeline.svg)
- **Expand: mimic.** Clone trajectories with small action and state noise to increase episode count.
- **Label and filter: VLM annotator.** Judge task success so failed episodes can be removed before training.
- **Augment, optional: Cosmos Transfer.** Re-render camera streams with varied lighting and backgrounds while preserving trajectories.
- **Convert: LeRobot.** Package the recording for GR00T or openpi and detect dimensions and camera keys.
- **Visualize.** Review camera videos and joint timelines before using GPU time for training.

*Optional deep dive:*

```text
Walk me through the data-prep tools (mimic, the VLM annotator, the LeRobot converter) and how each transforms an HDF5 recording while leaving actions and joint states intact.
```

## What’s Next?

You now have a labeled, converted, and inspected LeRobot dataset. In [Imitation Learning with GR00T](https://docs.nvidia.com/learning/physical-ai/getting-started-with-isaac-for-healthcare/latest/agentic-workflows/05-imitation-learning-gr00t.md) you’ll fine-tune a policy on it.