LeRobot v3 dataloader

[oxkitsune]

Related

• Closes Can't load lerobot v3.0 datasets #11678

Edit
New issue

What

• This PR is on top of Refactor LeRobot v2 dataloader #12066 *

This includes support for LeRobot datasets v3.0

[github-actions]

Web viewer built successfully.

Result	Commit	Link	Manifest
✅	17ec891	https://rerun.io/viewer/pr/12071	+nightly +main

View image diff on kitdiff.

^{Note: This comment is updated whenever you push a commit.}

[Copilot]

Pull request overview

This PR adds support for LeRobot v3 datasets to the Rerun dataloader. LeRobot v3 introduces changes to the dataset format, including the use of Parquet files for episode metadata and tasks (instead of JSONL files in v2), and feature-specific chunk/file indices for videos that allow multiple episodes to share video files more efficiently. The implementation follows the established patterns from the v2 dataloader.

Key Changes

• Implements v3 dataset loading with episode data caching for improved performance
• Adds video timestamp-based filtering to extract precise video segments per episode
• Includes feature-specific file metadata tracking for more flexible video/image organization

Reviewed changes

Copilot reviewed 5 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/python/release_checklist/check_lerobot_v3_dataloader.py	New test file for validating v3 dataset loading, consistent with v2 test structure
crates/store/re_data_loader/src/loader_lerobot.rs	Updated to support v3 datasets by removing the "unsupported" error and adding v3 loading function
crates/store/re_data_loader/src/lerobot/mod.rs	Updated module documentation to reflect v2 and v3 support
crates/store/re_data_loader/src/lerobot/datasetv3.rs	Complete v3 implementation with episode caching, video filtering, and Parquet-based metadata loading
crates/store/re_data_loader/Cargo.toml	Adds re_video dependency required for video processing
Cargo.lock	Updated lock file with new dependency

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Inconsistent indentation: the method call chain should be properly indented. The .sample_data_in_stream_format(&chunk) call on line 505 should align with the start of the chain, and video on line 504 should be indented to match standard Rust formatting conventions.

Suggested change

	.sample_data_in_stream_format(&chunk)
	.with_context(|| {
	format!(
	"Failed to convert sample {sample_idx} for feature '{observation}' to the expected codec stream format"
	)
	})?;
	.sample_data_in_stream_format(&chunk)
	.with_context(|| {
	format!(
	"Failed to convert sample {sample_idx} for feature '{observation}' to the expected codec stream format"
	)
	})?;

[MichaelGrupp]

Let's use pathlib instead of os.path

[MichaelGrupp]

Suggested change

	import os
	from argparse import Namespace
	from argparse import Namespace
	from pathlib import Path

[MichaelGrupp]

This f string interpolation looks unnecessary. Nicer modern way with pathlib is also:

Suggested change

	rec = rr.script_setup(args, f"{os.path.basename(__file__)}", recording_id="episode_0")
	rec = rr.script_setup(args, Path(__file__).name, recording_id="episode_0")

[MichaelGrupp]

Suggested change

	dataset_path = os.path.dirname(__file__) + "/.datasets/v30_apple_storage"
	dataset_path = Path(__file__).parent / ".datasets/v30_apple_storage"

[MichaelGrupp]

Suggested change

	rec = rr.script_setup(args, f"{os.path.basename(__file__)}", recording_id=f"episode_{i}")
	rec = rr.script_setup(args, Path(__file__).name, recording_id=f"episode_{i}")