fix(loader): detect LeRobot v3.0 datasets and guide users to bundled v3->v2 converter

[cagataycali]

What

LeRobotEpisodeLoader._load_metadata assumes a LeRobot v2.1 layout (single meta/episodes.jsonl). When pointed at a v3.0 dataset (e.g. lerobot/svla_so101_pickplace, BobShan/double_folding_towel_v3.0) the loader currently dies with a bare FileNotFoundError on meta/episodes.jsonl, because v3 keeps per-chunk episode metadata under meta/episodes/chunk-XXX/file-YYY.parquet instead.

This is a real footgun — the LeRobot team is steadily migrating canonical datasets to v3.0, so users following any GR00T-N1.5/N1.7 finetuning recipe with the latest dataset releases hit a 30-second crash with no hint at the cause.

How

Detect the v3.0 layout in two ways:

1. meta/episodes/ directory exists, or
2. info.json["codebase_version"] starts with v3.

When detected, raise a targeted FileNotFoundError that points the user at the converter already shipped in this repo (scripts/lerobot_conversion/convert_v3_to_v2.py), including the exact command to run.

Behaviour for v2.1 datasets is unchanged — the new code only fires when meta/episodes.jsonl is missing.

Diff

# gr00t/data/dataset/lerobot_episode_loader.py
episodes_path = meta_dir / LEROBOT_EPISODES_FILENAME
+ if not episodes_path.exists():
+     v3_episodes_dir = meta_dir / "episodes"
+     codebase_version = self.info_meta.get("codebase_version", "unknown")
+     if v3_episodes_dir.exists() or str(codebase_version).startswith("v3"):
+         raise FileNotFoundError(
+             f"... v3.0 layout detected, run convert_v3_to_v2.py ..."
+         )
+     raise FileNotFoundError(
+         f"{episodes_path} not found for dataset {self.dataset_path}. "
+         f"Expected a LeRobot v2.1 layout (meta/episodes.jsonl)."
+     )
with open(episodes_path, "r") as f:
    self.episodes_metadata = [json.loads(line) for line in f]

Repro of the original failure

hf download lerobot/svla_so101_pickplace --repo-type dataset --local-dir /tmp/svla
python -c "from gr00t.data.dataset.lerobot_episode_loader import LeRobotEpisodeLoader; \LeRobotEpisodeLoader('/tmp/svla', modality_configs={})"
# before: FileNotFoundError: [Errno 2] No such file or directory: '/tmp/svla/meta/episodes.jsonl'
# after:  FileNotFoundError: ... v3.0 layout, which is not yet natively supported.
#         Convert it with the helper shipped in this repo:
#             cd scripts/lerobot_conversion
#             uv venv && source .venv/bin/activate
#             uv pip install -e .
#             python convert_v3_to_v2.py --repo-id <hf-org/dataset-name>

Notes / future work

• This PR is intentionally minimal — it does not add native v3.0 support, just a clear pointer to the existing converter.
• A follow-up could call the converter automatically on load, but I'd rather keep that opt-in given v3 -> v2 conversion rewrites the dataset on disk.
• Tested locally on Linux aarch64 (Thor), Python 3.12.

cc @ ## Checklist

• No public API change for valid v2.1 datasets
• No new dependencies
• Syntactically validated (python -c "import ast; ast.parse(...)")
• Branch is rebased on main (3df8b38)