LeRobot Configuration

SUMMARY

Configure dataset reading with LeRobotDatasetReaderConfig, recording with LeRobotDatasetWriterConfig, and video output with dedicated RGB and depth encoder configurations.

Choose a Configuration

Configuration	Use
LeRobotDatasetReaderConfig	Loading, episode selection, temporal sampling, transforms, and decoding.
LeRobotDatasetWriterConfig	Recording, buffering, streaming encoding, and shard sizes.
RGBEncoderConfig	RGB video codec, quality, pixel format, and performance.
DepthEncoderConfig	Depth video codec, quality, pixel format, range, and logarithmic encoding.

Dataset Read Configuration

Use LeRobotDatasetReaderConfig when loading an existing dataset.

Parameter Configuration

Key	Type	Default	Description
root	str | Path | None	None	Local dataset root.
revision	str | None	None	Hub branch, tag, or commit.
tolerance_s	float	1e-4	Timestamp synchronization tolerance.
video_backend	str | None	None	Video decoder backend.
depth_output_unit	str	"mm"	Decoded depth unit, "m" or "mm".
episode_indices	list[int] | None	None	Episodes to load.
episode_filter	Callable | None	None	Metadata predicate applied to candidate episodes.
image_transforms	Callable | None	None	Transform applied after image decoding.
delta_timestamps	dict[str, list[float]] | None	None	Temporal offsets requested per feature.
force_cache_sync	bool	False	Refresh cached Hub data.
return_uint8	bool	False	Return visual values as uint8 rather than normalized tensors.
download_videos	bool	True	Download video assets when fetching from the Hub.
token	str | bool | None	None	Hugging Face authentication token behavior.

Example

from telekinesis.dataengine import datasets

config = datasets.LeRobotDatasetReaderConfig(
    episode_indices=[0, 1],
    delta_timestamps={"action": [0.0, 0.1, 0.2]},
    depth_output_unit="m",
)

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

Dataset Writer Configuration

Pass LeRobotDatasetWriterConfig to LeRobotDatasetLogger when recording episodes.

Parameter Configuration

Key	Type	Default	Description
tolerance_s	float	1e-4	Timestamp tolerance.
image_writer_processes	int	0	Async image-writer processes.
image_writer_threads	int	0	Async image-writer threads.
video_backend	str | None	None	Video backend.
batch_encoding_size	int	1	Episodes accumulated before video encoding.
rgb_encoder	RGBEncoderConfig | None	None	RGB codec configuration.
depth_encoder	DepthEncoderConfig | None	None	Depth codec and range configuration.
streaming_encoding	bool	False	Encode camera frames during capture.
encoder_queue_maxsize	int	30	Buffered frames per streaming camera.
encoder_threads	int | None	None	Global encoding thread count.
metadata_buffer_size	int	10	Episode metadata records buffered before flush.
max_video_files_size_in_mb	int | None	None	Maximum video shard size.
max_data_files_size_in_mb	int | None	None	Maximum Parquet shard size.

Example

from telekinesis.dataengine import data_loggers, datasets

write_config = datasets.LeRobotDatasetWriterConfig(
    tolerance_s=1e-4,
    streaming_encoding=True,
    encoder_queue_maxsize=60,
)

logger = data_loggers.LeRobotDatasetLogger(
    repo_id="user/my_dataset",
    local_path="results/user/my_dataset",
    mode="create",
    fps=30,
    features=features,
    config=write_config,
)

RGB Encoder Configuration

Use RGBEncoderConfig to control RGB video encoding.

INFO

Please note this in in active development. And the parameter will be updated freqeuntly.

Parameter Configuration

Key	Type	Default	Description
vcodec	str	"libsvtav1"	Video codec, or "auto" for automatic selection.
pix_fmt	str	"yuv420p"	Encoded pixel format.
g	int | None	2	Group-of-pictures size.
crf	int | float | None	30	Quality level; lower values generally increase quality and file size.
preset	int | str | None	Codec-specific	Encoding speed and quality preset.
fast_decode	int	0	Codec-specific fast-decoding tuning.
video_backend	str	"pyav"	Encoding backend.
extra_options	dict[str, Any]	{}	Additional codec options.

Example

from telekinesis.dataengine import datasets
from telekinesis.dataengine.datasets.lerobot.configs import RGBEncoderConfig

rgb_encoder = RGBEncoderConfig(
    vcodec="libsvtav1",
    pix_fmt="yuv420p",
    crf=24,
)

write_config = datasets.LeRobotDatasetWriterConfig(
    rgb_encoder=rgb_encoder,
)

Depth Encoder Configuration

Use DepthEncoderConfig to control depth video encoding and represented depth range.

INFO

Please note this in in active development. And the parameter will be updated freqeuntly.

Parameter Configuration

Key	Type	Default	Description
vcodec	str	"hevc"	Video codec, or "auto" for automatic selection.
pix_fmt	str	"gray12le"	Encoded pixel format.
g	int | None	2	Group-of-pictures size.
crf	int | float | None	30	Quality level; lower values generally increase quality and file size.
preset	int | str | None	Codec-specific	Encoding speed and quality preset.
fast_decode	int	0	Codec-specific fast-decoding tuning.
video_backend	str	"pyav"	Encoding backend.
extra_options	dict[str, Any]	{}	Additional codec options.
depth_min	float	0.01	Minimum represented depth.
depth_max	float	10.0	Maximum represented depth.
shift	float	3.5	Shift used by depth encoding.
use_log	bool	True	Encode depth values logarithmically.

Example

from telekinesis.dataengine import datasets
from telekinesis.dataengine.datasets.lerobot.configs import DepthEncoderConfig

depth_encoder = DepthEncoderConfig(
    vcodec="hevc",
    pix_fmt="gray12le",
    crf=24,
    depth_min=0.01,
    depth_max=10.0,
    use_log=True,
)

write_config = datasets.LeRobotDatasetWriterConfig(
    depth_encoder=depth_encoder,
)

Next Steps

Load with a Configuration

Select episodes, temporal windows, transforms, and decoding behavior.

Load dataset →

Create a Dataset

Define a frame rate and feature schema for a new writable dataset.

Create dataset →

Record Demonstrations

Apply writer and encoder settings during live episode recording.

Start recording →