LeRobot Dataset

SUMMARY

LeRobotDataset is the Telekinesis Data Engine interface for working with LeRobot v3 datasets. It supports creating, writing, loading, inspecting, and visualizing robot datasets containing synchronized observations, robot state, actions, and task annotations.

Install Telekinesis Skill Library

Create an API key, install telekinesis-ai and module-specific installations and dependencies in a Python environment.

Open installation →

Install

Install the LeRobot dependencies for the Data Engine:

pip install "telekinesis-ai[dataengine-lerobot]"

When to Use LeRobotDataset

Use LeRobotDataset when you want to:

• Create a LeRobot v3 dataset programmatically
• Write robot observations, actions, and task data into episodes
• Load an existing local or Hugging Face dataset
• Read frames for training or analysis
• Inspect dataset metadata and feature definitions
• Visualize recorded robot demonstrations
• Resume writing to an existing writable dataset

For live episode-based data collection, use LeRobotDatasetLogger instead.

What Does LeRobotDataset Provide?

LeRobotDataset provides the dataset layer for LeRobot v3 data, handling storage, metadata, frame access, and writable dataset operations through a single interface.

Capability	Description
Create	Create a new LeRobot dataset from a frame rate and feature schema.
Read	Access individual frames through the dataset interface.
Load	Open datasets from a local path or supported remote repository.
Write	Add frames and save episodes to a writable dataset.
Inspect	Read metadata, features, episode information, and dataset properties.
Visualize	Inspect recorded observations and actions visually.
Resume	Reopen an existing dataset for additional writes.

How a LeRobot Dataset Is Structured

A LeRobot dataset stores robot demonstrations as a collection of episodes.

Each episode contains a time-ordered sequence of frames.

Each frame follows the dataset's feature schema and can contain synchronized values such as:

• camera observations
• robot state
• actions
• additional sensor observations
• task information

Conceptually, the dataset is organized as:

LeRobot Dataset
│
├── Episode 0
│   ├── Frame 0
│   ├── Frame 1
│   ├── ...
│   └── Frame N
│
├── Episode 1
│   ├── Frame 0
│   ├── Frame 1
│   └── ...
│
└── Episode N

The feature schema defines what every frame contains, including each feature's data type, shape, and dimension names.

{
    "observation.camera_rgb": {
        "dtype": "video",
        "shape": [3, 64, 64],
        "names": ["channels", "height", "width"],
    },
    "observation.state": {
        "dtype": "float32",
        "shape": [7],
        "names": [
            "joint_1",
            "joint_2",
            "joint_3",
            "joint_4",
            "joint_5",
            "joint_6",
            "gripper_state",
        ],
    },
    "action": {
        "dtype": "float32",
        "shape": [7],
        "names": [
            "joint_1",
            "joint_2",
            "joint_3",
            "joint_4",
            "joint_5",
            "joint_6",
            "gripper",
        ],
    },
}

Storage Layout

On disk, LeRobot separates frame data, visual streams, and dataset metadata:

dataset_name/
├── .cache/
├── data/
│   └── chunk-000/
├── meta/
│   ├── episodes/
│   │   └── chunk-000/
│   │       ├── file-000.parquet
|   |       ├── ... 
│   │       └── file-00n.parquet
│   ├── info.json
│   ├── stats.json
│   └── tasks.parquet
├── videos/
│   ├── observation.images.camera1/
│   │   └── chunk-000/
|   |   ├── ... 
│   └── observation.images.camera3/
│       └── chunk-000/
├── .gitattributes
└── README.md

Location	Purpose
data/	Stores frame-level non-visual data in Parquet shards.
videos/	Stores encoded visual observations, organized by camera feature.
meta/info.json	Stores dataset-level information such as the feature schema, frame rate, and storage layout.
meta/stats.json	Stores dataset statistics used for inspection and normalization.
meta/episodes/	Stores per-episode metadata such as episode lengths and task associations.
meta/tasks.parquet	Stores task descriptions and their dataset mappings.

Guides

Use the guides below depending on whether you want to create, write, load, inspect, or visualize a LeRobot dataset.

Create a LeRobot Dataset

Create a new writable LeRobot dataset from a frame rate, robot type, and feature schema.

Create dataset →

Write to a LeRobot Dataset

Add frames, save episodes, and finalize a writable LeRobot dataset.

Write data →

Load a LeRobot Dataset

Open an existing local or remote LeRobot dataset and access its frames.

Load dataset →

Inspect Dataset Metadata

Inspect feature definitions, frame rate, robot type, episodes, and other dataset metadata.

Inspect metadata →

Visualize a LeRobot Dataset

Visualize recorded observations and actions to inspect the contents of a dataset.

Visualize dataset →