lerobot-edit-dataset

Edit, transform, and manage LeRobot datasets

The lerobot-edit-dataset command provides tools for modifying datasets including deleting episodes, splitting, merging, removing features, modifying tasks, and converting image datasets to video format.

Command

lerobot-edit-dataset [OPTIONS]

Location: src/lerobot/scripts/lerobot_edit_dataset.py

Overview

Supported operations:

• delete_episodes: Remove specific episodes from dataset
• split: Split dataset into train/val/test sets
• merge: Combine multiple datasets into one
• remove_feature: Delete camera or sensor features
• modify_tasks: Update task descriptions
• convert_image_to_video: Convert image dataset to video format
• info: Display dataset statistics

Key Options

Dataset Options

--repo_id · str

Input dataset repository ID. Required for all operations except merge.

--root · str

Local path to input dataset. Defaults to $HF_LEROBOT_HOME/{repo_id}.

--new_repo_id · str

Output dataset repository ID. Required for merge, optional for others.

--new_root · str

Local path for output dataset. Defaults to $HF_LEROBOT_HOME/{new_repo_id}.

--push_to_hub · bool · default: "false"

Upload result to Hugging Face Hub.

Operation Configuration

--operation.type · str · required

Operation to perform: delete_episodes, split, merge, remove_feature, modify_tasks, convert_image_to_video, or info.

Additional options depend on operation type (see examples below).

Usage Examples

Delete Episodes

Remove specific episodes from a dataset:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=delete_episodes \
  --operation.episode_indices="[0, 2, 5]"

This creates a backup at pusht_old and modifies the dataset in-place.

Save to New Dataset

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --new_repo_id=lerobot/pusht_filtered \
  --operation.type=delete_episodes \
  --operation.episode_indices="[0, 2, 5, 10, 15]"

Split Dataset

Split into train/val sets by fractions:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=split \
  --operation.splits='{"train": 0.8, "val": 0.2}'

Creates:

• lerobot/pusht_train (80% of episodes)
• lerobot/pusht_val (20% of episodes)

Split by Episode Indices

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=split \
  --operation.splits='{"train": [0, 1, 2, 3, 4], "val": [5, 6], "test": [7, 8, 9]}'

Custom Output Location

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --new_root=/path/to/splits \
  --operation.type=split \
  --operation.splits='{"train": 0.7, "val": 0.15, "test": 0.15}'

Creates splits in /path/to/splits/train, /path/to/splits/val, /path/to/splits/test.

Merge Datasets

Combine multiple datasets:

lerobot-edit-dataset \
  --new_repo_id=lerobot/pusht_merged \
  --operation.type=merge \
  --operation.repo_ids="['lerobot/pusht_train', 'lerobot/pusht_val']"

Merge from Local Paths

lerobot-edit-dataset \
  --new_repo_id=lerobot/combined_dataset \
  --operation.type=merge \
  --operation.repo_ids="['dataset1', 'dataset2']" \
  --operation.roots="['/path/to/dataset1', '/path/to/dataset2']"

Merge and Push to Hub

lerobot-edit-dataset \
  --new_repo_id=myuser/merged_dataset \
  --operation.type=merge \
  --operation.repo_ids="['lerobot/pusht', 'lerobot/aloha_mobile_cabinet']" \
  --push_to_hub=true

Remove Feature

Delete camera or sensor features:

lerobot-edit-dataset \
  --repo_id=lerobot/aloha_mobile_cabinet \
  --operation.type=remove_feature \
  --operation.feature_names="['observation.images.top']"

Removes the observation.images.top camera from the dataset.

Remove Multiple Features

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --new_repo_id=lerobot/pusht_minimal \
  --operation.type=remove_feature \
  --operation.feature_names="['observation.images.wrist', 'observation.depth']"

Modify Tasks

WARNING: Modifies dataset in-place! Back up first.

Set Single Task for All Episodes

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=modify_tasks \
  --operation.new_task="Push the T-shaped block to the target"

Set Different Tasks per Episode

lerobot-edit-dataset \
  --repo_id=myuser/my_dataset \
  --operation.type=modify_tasks \
  --operation.episode_tasks='{"0": "Pick red cube", "1": "Pick blue cube", "2": "Pick green cube"}'

Set Default with Overrides

lerobot-edit-dataset \
  --repo_id=myuser/my_dataset \
  --operation.type=modify_tasks \
  --operation.new_task="Standard pick and place" \
  --operation.episode_tasks='{"5": "Pick fragile object", "10": "Pick heavy object"}'

Convert Image to Video

Convert image-based dataset to video format for smaller storage:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht_image \
  --new_repo_id=lerobot/pusht_video \
  --operation.type=convert_image_to_video

Custom Video Settings

lerobot-edit-dataset \
  --repo_id=lerobot/pusht_image \
  --new_repo_id=lerobot/pusht_video \
  --operation.type=convert_image_to_video \
  --operation.vcodec=h264_nvenc \
  --operation.crf=23 \
  --operation.num_workers=8

Convert and Push to Hub

lerobot-edit-dataset \
  --repo_id=myuser/dataset_image \
  --new_repo_id=myuser/dataset_video \
  --operation.type=convert_image_to_video \
  --push_to_hub=true

Dataset Info

Display dataset statistics:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=info

Output:

======Info lerobot/pusht
Repository ID: lerobot/pusht 
Total episode: 100 
Total task: 1 
Total frame(Actual Count): 4500(4500) 
Average frame per episode: 45.0
Average episode time(sec): 1.5
FPS: 30
Size: 1234.5 MB

Show Feature Details

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=info \
  --operation.show_features=true

Shows full feature schema in JSON format.

Video Encoding Options

For convert_image_to_video operation:

--operation.vcodec · str · default: "libsvtav1"

Video codec: h264, hevc, libsvtav1, h264_nvenc, h264_videotoolbox, etc.

--operation.pix_fmt · str · default: "yuv420p"

Pixel format for compatibility.

--operation.g · int · default: "2"

GOP size (keyframe interval).

--operation.crf · int · default: "30"

Constant rate factor (lower = higher quality, larger file).

--operation.num_workers · int · default: "4"

Number of parallel encoding workers.

In-Place vs New Dataset

In-Place Modification

When new_repo_id and new_root are not specified:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --operation.type=delete_episodes \
  --operation.episode_indices="[0]"

• Creates backup: pusht_old/
• Modifies original dataset location
• Use for permanent changes

New Dataset Creation

When new_repo_id or new_root is specified:

lerobot-edit-dataset \
  --repo_id=lerobot/pusht \
  --new_repo_id=lerobot/pusht_modified \
  --operation.type=delete_episodes \
  --operation.episode_indices="[0]"

• Original dataset unchanged
• Creates new dataset at new_root or $HF_LEROBOT_HOME/new_repo_id
• Use when preserving original

Path Semantics

Default Paths

# Uses $HF_LEROBOT_HOME/lerobot/pusht
lerobot-edit-dataset --repo_id=lerobot/pusht ...

# Uses $HF_LEROBOT_HOME/lerobot/pusht_new
lerobot-edit-dataset --repo_id=lerobot/pusht --new_repo_id=lerobot/pusht_new ...

Explicit Paths

# Uses /path/to/dataset
lerobot-edit-dataset --repo_id=lerobot/pusht --root=/path/to/dataset ...

# Uses /path/to/output
lerobot-edit-dataset --repo_id=lerobot/pusht --new_root=/path/to/output ...

Programmatic Usage

from lerobot.scripts.lerobot_edit_dataset import (
    edit_dataset,
    EditDatasetConfig,
    DeleteEpisodesConfig
)

config = EditDatasetConfig(
    repo_id="lerobot/pusht",
    new_repo_id="lerobot/pusht_filtered",
    operation=DeleteEpisodesConfig(
        episode_indices=[0, 2, 5]
    ),
    push_to_hub=False,
)

edit_dataset(config)

Best Practices

1. Backup First: Always backup before in-place modifications
2. Test on Subset: Test operations on small datasets first
3. Verify Output: Check result before pushing to Hub
4. Use New Dataset: Prefer new_repo_id over in-place for safety
5. Document Changes: Record what operations were performed

See Also

• LeRobotDataset - Dataset format and structure
• lerobot-record - Record new datasets
• lerobot-dataset-viz - Visualize datasets