Fix: lerobot-dataset-edit merge with custom root paths locally

[riochuong]

Title

fix(scripts): handle custom root paths in dataset merge operation

CREDIT:

Fix and tests are written with assistant from Claude 4.5 Sonnet. I did verify the changes and executing testing locally to make sure it works as expected

Type / Scope

• Type: Bug
• Scope: scripts/lerobot_edit_dataset, datasets/dataset_tools

Summary / Motivation

The lerobot-dataset-edit --operation.type merge command failed when users specified custom --root paths for merging local datasets. The handle_merge() function was passing the root directory directly to LeRobotDataset() without appending the individual dataset's repo_id, causing the loader to look for metadata files in the wrong location (e.g., /path/to/datasets/meta/info.json instead of /path/to/datasets/dataset1/meta/info.json).

This fix enables users to merge locally stored datasets that are organized in custom directories, which is a common workflow when working with self-collected robotic datasets before uploading to HuggingFace Hub.

Related issues

• Fixes: N/A (discovered during local dataset management workflow)
• Related: N/A

What changed

Code changes:

• src/lerobot/scripts/lerobot_edit_dataset.py (lines 248-250): Updated handle_merge() to construct full dataset paths by appending repo_id to custom root when provided:

  Before (buggy):

  datasets = [LeRobotDataset(repo_id, root=cfg.root) for repo_id in cfg.operation.repo_ids]

  After (fixed):

  datasets = [
  LeRobotDataset(repo_id, root=Path(cfg.root) / repo_id if cfg.root else None)
  for repo_id in cfg.operation.repo_ids
  ]
  Test additions:

• tests/datasets/test_dataset_tools.py: Added 3 comprehensive test functions:

  • test_handle_merge_with_custom_root() - Validates the bug fix with custom root paths
  • test_handle_merge_without_custom_root() - Ensures default behavior still works
  • test_handle_merge_custom_root_preserves_metadata() - Verifies metadata preservation during merge

Breaking changes: None. This is a pure bug fix that maintains backward compatibility.

How was this tested

Tests added:

• test_handle_merge_with_custom_root - Creates two datasets in a custom root directory, merges them, and verifies the merged dataset is created in the correct location with proper episode counts.
• test_handle_merge_without_custom_root - Tests that the default behavior (no custom root) continues to work correctly.
• test_handle_merge_custom_root_preserves_metadata - Ensures that dataset metadata (FPS, features, episode counts, frame counts) are correctly preserved after merging with custom roots.

Manual testing:

• Successfully merged two local datasets with custom root paths using:
  lerobot-dataset-edit
  --repo_id merged_dataset
  --root /path/to/datasets
  --operation.type merge
  --operation.repo_ids "['dataset1', 'dataset2']"
  Test results:
  $ pytest tests/datasets/test_dataset_tools.py -k merge -v

All 7 merge tests pass (3 new + 4 existing)## How to run locally (reviewer)

Run all merge-related tests:
pytest tests/datasets/test_dataset_tools.py -k merge -vRun only the new tests:
pytest tests/datasets/test_dataset_tools.py::test_handle_merge_with_custom_root -v
pytest tests/datasets/test_dataset_tools.py::test_handle_merge_without_custom_root -v
pytest tests/datasets/test_dataset_tools.py::test_handle_merge_custom_root_preserves_metadata -vManual test with local datasets:

Create two test datasets in a custom directory

lerobot-record --some-config # or use existing datasets

Try merging with custom root (this would have failed before the fix)

lerobot-dataset-edit
--repo_id test_merged
--root /path/to/your/datasets
--operation.type merge
--operation.repo_ids "['dataset1', 'dataset2']"## Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated (N/A - internal bug fix, no user-facing API changes)
• CI is green

Reviewer notes

Focus areas:

• Line 248-250 in lerobot_edit_dataset.py: Verify the path construction logic correctly handles both None and custom root cases
• Test coverage: The three new tests cover the bug scenario (custom root), default behavior (no root), and metadata preservation
• Backward compatibility: Existing 4 merge tests still pass, confirming no regressions

Design note:
The fix follows the pattern already established by LeRobotDataset.__init__() which accepts either:

• root=None → uses HF_LEROBOT_HOME / repo_id
• root=Path(...) → uses the provided path as-is

By constructing Path(cfg.root) / repo_id before passing to LeRobotDataset(), we ensure the dataset loader receives the complete path to each individual dataset directory.

Edge cases covered:

• ✅ Custom root with multiple datasets
• ✅ No custom root (default HF_LEROBOT_HOME behavior)
• ✅ Metadata preservation (FPS, features, episode/frame counts)
• ✅ Different numbers of episodes per dataset

[sotanakamura]

As defined in LeRobotDataset, datasets will be stored under root/repo_id, so we need to standardize the dataset location in this way.

related to #2316

[riochuong]

does that mean for merge to work correctly on custom local folder the only way is to use root=None and move data to default HF_LEROBOT_HOME (not too bad but need to remember to set this as venv) ??