Load a LeRobot Dataset

SUMMARY

Use LeRobotDataset to open an existing LeRobot dataset from a local path or the Hugging Face Hub.

Loading can be customized with LeRobotDatasetConfig to select episodes, apply temporal sampling, configure transforms, and control decoding behavior.

Learn about LeRobot datasets

Review the dataset structure, feature schema, storage layout, and workflow.

View overview →

Method

Create a LeRobotDataset with a repository ID and an optional local path:

from telekinesis.dataengine import datasets

dataset = datasets.LeRobotDataset(
    repo_id="lerobot/pusht",
    local_path="results/lerobot/pusht",
)

If the dataset is not already available locally, it is resolved from the configured repository source and stored at the local path.

Once loaded, the dataset can be indexed directly:

frame = dataset[0]

Each item is returned as a frame dictionary containing the features defined by the dataset schema.

Parameter Configuration

Parameter	Type	Default	Description
repo_id	str	required	Dataset repository identifier.
local_path	str | Path | None	None	Local directory used to load or store the dataset.
config	LeRobotDatasetConfig | None	None	Loading, filtering, decoding, and transform configuration.

Config Reference

Use LeRobotDatasetConfig to control how the dataset is read:

config = datasets.LeRobotDatasetConfig(
    episode_indices=[0, 1, 2],
)

dataset = datasets.LeRobotDataset(
    repo_id="lerobot/pusht",
    local_path="results/lerobot/pusht",
    config=config,
)

For example, episode_indices limits the loaded dataset to a subset of episodes.

Other configuration options control temporal windows, transforms, video decoding, depth units, and cache behavior. For full confifuration details see Configuration.

Attribute Reference

Attribute	Type	Description
repo_id	str	Repository identifier used to load the dataset.
local_path	Path	Resolved local directory containing the dataset.
episode_indices	list[int] | None	Episode indices selected by the loading configuration. None means all episodes.
num_episodes	int	Number of loaded or selected episodes.
num_frames	int	Number of frames across the loaded or selected episodes.
features	dict[str, dict]	Feature schema describing the observations, actions, and other dataset values.

These can be used to verify which data was loaded before training, visualization, or further processing.

Indexing Reference

Input	Return Type	Description
dataset[index]	dict	Returns one decoded frame.
dataset[start:stop]	list[dict]	Returns a list of decoded frames.

Example

"""Example script demonstrating how to load a LeRobot dataset using the Telekinesis Data Engine."""

from pathlib import Path

from loguru import logger

from telekinesis.dataengine import datasets


def load_lerobot_dataset_example():
    """Load selected episodes from a LeRobot dataset."""

    # 1. Define the dataset identity and local storage path.
    repo_id = "lerobot/pusht"

    local_path = (
        Path(__file__).resolve().parent.parent.parent.parent
        / "results"
        / repo_id
    )

    # 2. Configure how the dataset should be loaded.
    config = datasets.LeRobotDatasetConfig(
        episode_indices=[0, 1, 2],
    )

    # 3. Load the dataset.
    dataset = datasets.LeRobotDataset(
        repo_id=repo_id,
        local_path=local_path,
        config=config,
    )

    logger.info("LeRobot dataset loaded successfully.")
    logger.info(dataset)

    logger.info("Selected episodes: {}", dataset.episode_indices)
    logger.info("Number of selected episodes: {}", dataset.num_episodes)
    logger.info("Number of selected frames: {}", dataset.num_frames)

    # 4. Access the first frame.
    first_frame = dataset[0]

    logger.info("First frame: {}", first_frame)


if __name__ == "__main__":
    load_lerobot_dataset_example()

Next Steps

Inspect Dataset Metadata

Examine the dataset schema, frame rate, episodes, statistics, and storage metadata.

Inspect metadata →

Visualize a LeRobot Dataset

Inspect recorded observations and actions in Rerun.

Visualize dataset →

Configure Dataset Loading

Configure filtering, temporal sampling, transforms, depth units, and decoding behavior.

View configuration →