Support streaming from HF Storage Buckets in StreamingLeRobotDataset

Add an opt-in repo_type="bucket" parameter to StreamingLeRobotDataset
and LeRobotDatasetMetadata so a dataset can be streamed directly from an
HF Storage Bucket (hf://buckets/...) with no local download.

When repo_type="bucket":
- skip git-version resolution (buckets have no refs/tags),
- pull the meta/ directory via HfFileSystem.get,
- point url_root at hf://buckets/{repo_id},
- read parquet shards via load_dataset("parquet",
  data_files="hf://buckets/{repo_id}/data/*/*.parquet", ...).

The default repo_type="dataset" preserves all existing behavior.
LeRobotDataset (non-streaming) and create() are unchanged.

Closes #3969

Signed-off-by: Sundar Raghavan <sdraghav@amazon.com>

src/lerobot/datasets/dataset_metadata.py

@@ -73,6 +73,7 @@ class LeRobotDatasetMetadata:
         revision: str | None = None,
         force_cache_sync: bool = False,
         metadata_buffer_size: int = 10,
+        repo_type: str = "dataset",
         *,
         token: str | bool | None = None,
     ):
@@ -96,12 +97,15 @@ class LeRobotDatasetMetadata:
                 even when local files exist.
             metadata_buffer_size: Number of episode metadata records to buffer
                 in memory before flushing to parquet.
+            repo_type: Repository type: "dataset" (default) or "bucket" for an
+                HF Storage Bucket streamed over hf://buckets/.
             token: Authentication token used for Hub requests. Pass a string
                 token, ``True`` to require the locally stored token, ``False``
                 to disable authentication, or ``None`` to use the Hugging Face
                 Hub default.
         """
         self.repo_id = repo_id
+        self.repo_type = repo_type
         self.revision = revision if revision else CODEBASE_VERSION
         self._requested_root = Path(root) if root is not None else None
         self.root = self._requested_root if self._requested_root is not None else HF_LEROBOT_HOME / repo_id
@@ -118,7 +122,7 @@ class LeRobotDatasetMetadata:
                 raise FileNotFoundError
             self._load_metadata()
         except (FileNotFoundError, NotADirectoryError):
-            if is_valid_version(self.revision):
+            if self.repo_type != "bucket" and is_valid_version(self.revision):
                 if token is None:
                     self.revision = get_safe_version(self.repo_id, self.revision)
                 else:
@@ -232,6 +236,22 @@ class LeRobotDatasetMetadata:
         *,
         token: str | bool | None = None,
     ) -> None:
+        if getattr(self, "repo_type", "dataset") == "bucket":
+            from huggingface_hub import HfFileSystem
+
+            fs = HfFileSystem()
+            dest = (
+                self._requested_root
+                if self._requested_root is not None
+                else (HF_LEROBOT_HUB_CACHE / ("buckets--" + self.repo_id.replace("/", "--")))
+            )
+            dest = Path(dest)
+            dest.mkdir(parents=True, exist_ok=True)
+            # fs.get copies the SOURCE dir INTO the target, so target=dest (not
+            # dest/meta) lands the tree as dest/meta/... rather than dest/meta/meta.
+            fs.get(f"hf://buckets/{self.repo_id}/meta", str(dest), recursive=True)
+            self.root = dest
+            return
         token_kwargs = {} if token is None else {"token": token}
         if self._requested_root is None:
             self.root = Path(
@@ -262,6 +282,8 @@ class LeRobotDatasetMetadata:
     @property
     def url_root(self) -> str:
         """Hugging Face Hub URL root for this dataset."""
+        if getattr(self, "repo_type", "dataset") == "bucket":
+            return f"hf://buckets/{self.repo_id}"
         return f"hf://datasets/{self.repo_id}"
 
     @property

src/lerobot/datasets/streaming_dataset.py

@@ -242,6 +242,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         self,
         repo_id: str,
         root: str | Path | None = None,
+        repo_type: str = "dataset",
         episodes: list[int] | None = None,
         image_transforms: Callable | None = None,
         delta_timestamps: dict[list[float]] | None = None,
@@ -266,6 +267,8 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             root (Path | None, optional): Local directory to use for local datasets. When omitted, Hub
                 metadata is resolved through a revision-safe snapshot cache under
                 ``$HF_LEROBOT_HOME/hub``.
+            repo_type (str, optional): "dataset" (default) or "bucket" to stream
+                from an HF Storage Bucket over hf://buckets/.
             episodes (list[int] | None, optional): If specified, this will only load episodes specified by
                 their episode_index in this list.
             image_transforms (Callable | None, optional): Transform to apply to image data.
@@ -288,6 +291,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         """
         super().__init__()
         self.repo_id = repo_id
+        self.repo_type = repo_type
         self._requested_root = Path(root) if root else None
         self.root = self._requested_root if self._requested_root is not None else HF_LEROBOT_HOME / repo_id
         self.streaming_from_local = root is not None
@@ -317,6 +321,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             self._requested_root,
             self.revision,
             force_cache_sync=force_cache_sync,
+            repo_type=self.repo_type,
             token=token,
         )
         self.root = self.meta.root
@@ -345,6 +350,14 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             self.delta_timestamps = delta_timestamps
             self.delta_indices = get_delta_indices(self.delta_timestamps, self.fps)
 
+        if self.repo_type == "bucket":
+            self.hf_dataset: datasets.IterableDataset = load_dataset(
+                "parquet",
+                data_files=f"hf://buckets/{self.repo_id}/data/*/*.parquet",
+                split="train",
+                streaming=self.streaming,
+            )
+        else:
             token_kwargs = {} if token is None or self.streaming_from_local else {"token": token}
             self.hf_dataset: datasets.IterableDataset = load_dataset(
                 self.repo_id if not self.streaming_from_local else str(self.root),

tests/datasets/test_streaming.py

@@ -14,7 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from types import SimpleNamespace
-from unittest.mock import Mock
+from unittest.mock import MagicMock, Mock, patch
 
 import numpy as np
 import pytest
@@ -23,6 +23,7 @@ import torch
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 
 import lerobot.datasets.streaming_dataset as streaming_dataset_module
+from lerobot.datasets.dataset_metadata import LeRobotDatasetMetadata
 from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
 from lerobot.datasets.utils import safe_shard
 from lerobot.utils.constants import ACTION
@@ -100,6 +101,7 @@ def test_streaming_dataset_forwards_hub_token_only_for_remote_data(tmp_path, mon
         requested_root,
         streaming_dataset_module.CODEBASE_VERSION,
         force_cache_sync=False,
+        repo_type="dataset",
         token=token,
     )
     if from_local:
@@ -430,3 +432,79 @@ def test_frames_with_delta_consistency_with_shards(
         assert all(t[1] for t in key_checks), (
             f"Checking {list(filter(lambda t: not t[1], key_checks))[0][0]} left and right were found different (i: {i}, frame_idx: {frame_idx})"
         )
+
+
+class _StopConstructionError(Exception):
+    """Sentinel raised from a patched load_dataset to halt __init__ after the branch under test."""
+
+
+def _fake_meta(*args, **kwargs):
+    """Minimal LeRobotDatasetMetadata stand-in exposing only what __init__ reads."""
+    meta = type("_Meta", (), {})()
+    meta.root = kwargs.get("root") or "/tmp/_streaming_meta"
+    meta.revision = kwargs.get("revision") or "v0"
+    meta._version = "v3.0"
+    meta.depth_keys = []
+    meta.image_keys = []
+    meta.rescale_depth_stats = lambda *_a, **_k: None
+    meta.repo_type = kwargs.get("repo_type", "dataset")
+    return meta
+
+
+@pytest.mark.parametrize(
+    "repo_type, expected_source, expected_data_files",
+    [
+        ("bucket", "parquet", "hf://buckets/{repo_id}/data/*/*.parquet"),
+        ("dataset", "{repo_id}", "data/*/*.parquet"),
+    ],
+)
+def test_streaming_repo_type_routes_load_dataset(repo_type, expected_source, expected_data_files):
+    """repo_type='bucket' loads parquet from hf://buckets/...; 'dataset' keeps the Hub-repo path."""
+    captured = {}
+
+    def fake_load_dataset(source, **kwargs):
+        captured["source"] = source
+        captured["data_files"] = kwargs.get("data_files")
+        raise _StopConstructionError
+
+    with (
+        patch("lerobot.datasets.streaming_dataset.LeRobotDatasetMetadata", _fake_meta),
+        patch("lerobot.datasets.streaming_dataset.check_version_compatibility", lambda *a, **k: None),
+        patch("lerobot.datasets.streaming_dataset.load_dataset", fake_load_dataset),
+        pytest.raises(_StopConstructionError),
+    ):
+        StreamingLeRobotDataset(DUMMY_REPO_ID, repo_type=repo_type)
+
+    assert captured["source"] == expected_source.format(repo_id=DUMMY_REPO_ID)
+    assert captured["data_files"] == expected_data_files.format(repo_id=DUMMY_REPO_ID)
+
+
+def test_bucket_metadata_url_root(tmp_path):
+    """repo_type='bucket' produces url_root pointing at hf://buckets/..."""
+    mock_fs = MagicMock()
+    with (
+        patch("huggingface_hub.HfFileSystem", return_value=mock_fs),
+        patch.object(LeRobotDatasetMetadata, "_load_metadata"),
+    ):
+        meta = LeRobotDatasetMetadata(
+            DUMMY_REPO_ID,
+            root=tmp_path,
+            repo_type="bucket",
+        )
+    assert meta.url_root == f"hf://buckets/{DUMMY_REPO_ID}"
+
+
+def test_bucket_skips_get_safe_version(tmp_path):
+    """repo_type='bucket' must NOT call get_safe_version (buckets have no git refs)."""
+    mock_fs = MagicMock()
+    with (
+        patch("lerobot.datasets.dataset_metadata.get_safe_version") as mock_gsv,
+        patch("huggingface_hub.HfFileSystem", return_value=mock_fs),
+        patch.object(LeRobotDatasetMetadata, "_load_metadata"),
+    ):
+        LeRobotDatasetMetadata(
+            DUMMY_REPO_ID,
+            root=tmp_path,
+            repo_type="bucket",
+        )
+    mock_gsv.assert_not_called()