feat(eval): optionally push recorded eval datasets to the Hub

- Extract _infer_shape_from_obs() to reduce nesting in feature conversion
- Move dataset creation into rollout() using its own env.reset() observation,
  eliminating the extra reset in run_one()
- Replace deepcopy with _shallow_copy_obs() for raw observation stashing
- Support batch_size > 1: each parallel env records to its own dataset
  (single env skips the env_0/ nesting for simplicity)
- One-time warning for env_features keys missing from observations
- Pass recording_dir + env_features through the call chain instead of
  a pre-built recording_dataset object

src/lerobot/annotations/steerable_pipeline/writer.py

@@ -54,6 +54,7 @@ from typing import Any
 import pyarrow as pa
 import pyarrow.parquet as pq
 
+from lerobot.datasets.io_utils import write_table_one_row_group_per_episode
 from lerobot.datasets.language import (
     EVENT_ONLY_STYLES,
     LANGUAGE_EVENTS,
@@ -274,12 +275,11 @@ class LanguageColumnsWriter:
         new_table = self._materialize_table(
             table, per_row_persistent, per_row_events, drop_old=self.drop_existing_subtask_index
         )
-        # Atomic replace: write to a sibling tmp path and rename so a crash
-        # mid-write can't leave a half-written shard that ``pq.read_table``
-        # would then fail to open. ``Path.replace`` is atomic on POSIX +
-        # Windows when source and target sit on the same filesystem.
+        # Re-emit one row group per episode (a bulk pq.write_table would collapse
+        # them into one). Write to a sibling tmp path and atomically rename so a
+        # crash mid-write can't leave a half-written shard.
         tmp_path = path.with_suffix(path.suffix + ".tmp")
-        pq.write_table(new_table, tmp_path)
+        write_table_one_row_group_per_episode(new_table, tmp_path)
         tmp_path.replace(path)
 
     def _materialize_table(

src/lerobot/configs/default.py

@@ -73,8 +73,17 @@ class EvalConfig:
     # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
     # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
     use_async_envs: bool = True
+    # Whether to record eval rollouts as a LeRobot dataset on disk.
+    recording: bool = False
+    # If set, push recorded eval datasets to the Hub under this repo id (one repo per task,
+    # suffixed by task and env index). Requires recording=true.
+    recording_repo_id: str | None = None
+    # Whether the pushed recording repositories should be private.
+    recording_private: bool = False
 
     def __post_init__(self) -> None:
+        if self.recording_repo_id is not None and not self.recording:
+            raise ValueError("eval.recording_repo_id requires eval.recording=true.")
         if self.batch_size == 0:
             self.batch_size = self._auto_batch_size()
         if self.batch_size > self.n_episodes:

src/lerobot/datasets/aggregate.py

@@ -32,6 +32,7 @@ from .feature_utils import features_equal_for_merge, get_hf_features_from_featur
 from .io_utils import (
     get_file_size_in_mb,
     get_parquet_file_size_in_mb,
+    to_parquet_one_row_group_per_episode,
     to_parquet_with_hf_images,
     write_info,
     write_stats,
@@ -551,6 +552,7 @@ def aggregate_data(src_meta, dst_meta, data_idx, data_files_size_in_mb, chunk_si
             aggr_root=dst_meta.root,
             hf_features=hf_features,
             concatenate=concatenate_data,
+            one_row_group_per_episode=True,
         )
 
         # Record the mapping from source to actual destination
@@ -628,6 +630,7 @@ def append_or_create_parquet_file(
     aggr_root: Path = None,
     hf_features: datasets.Features | None = None,
     concatenate: bool = True,
+    one_row_group_per_episode: bool = False,
 ) -> tuple[dict[str, int], tuple[int, int]]:
     """Appends data to an existing parquet file or creates a new one based on size constraints.
 
@@ -645,6 +648,8 @@ def append_or_create_parquet_file(
         aggr_root: Root path for the aggregated dataset.
         hf_features: Optional HuggingFace Features schema for proper image typing.
         concatenate: When False, always rotate to a new file instead of appending to the current one.
+        one_row_group_per_episode: True for DATA parquet (emit one row group per episode); False for
+            the episodes-metadata parquet (already one row per episode).
 
     Returns:
         tuple: (updated_idx, (dst_chunk, dst_file)) where updated_idx is the index dict
@@ -657,6 +662,8 @@ def append_or_create_parquet_file(
         dst_path.parent.mkdir(parents=True, exist_ok=True)
         if contains_images:
             to_parquet_with_hf_images(df, dst_path, features=hf_features)
+        elif one_row_group_per_episode:
+            to_parquet_one_row_group_per_episode(df, dst_path)
         else:
             df.to_parquet(dst_path)
         return idx, (dst_chunk, dst_file)
@@ -683,6 +690,8 @@ def append_or_create_parquet_file(
 
     if contains_images:
         to_parquet_with_hf_images(final_df, target_path, features=hf_features)
+    elif one_row_group_per_episode:
+        to_parquet_one_row_group_per_episode(final_df, target_path)
     else:
         final_df.to_parquet(target_path)
 

src/lerobot/datasets/dataset_metadata.py

@@ -15,6 +15,7 @@
 # limitations under the License.
 import contextlib
 from collections.abc import Callable
+from copy import deepcopy
 from pathlib import Path
 
 import numpy as np
@@ -709,7 +710,7 @@ class LeRobotDatasetMetadata:
 
         obj.root.mkdir(parents=True, exist_ok=False)
 
-        features = {**features, **DEFAULT_FEATURES}
+        features = {**deepcopy(features), **DEFAULT_FEATURES}
         _validate_feature_names(features)
 
         obj.tasks = None

src/lerobot/datasets/dataset_tools.py

@@ -27,6 +27,7 @@ import logging
 import shutil
 from collections.abc import Callable
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor, as_completed
+from copy import deepcopy
 from pathlib import Path
 
 import datasets
@@ -1101,7 +1102,9 @@ def _copy_episodes_metadata_and_stats(
     if dst_meta.video_keys and src_dataset.meta.video_keys:
         for key in dst_meta.video_keys:
             if key in src_dataset.meta.features:
-                dst_meta.info.features[key]["info"] = src_dataset.meta.info.features[key].get("info", {})
+                dst_meta.info.features[key]["info"] = deepcopy(
+                    src_dataset.meta.info.features[key].get("info", {})
+                )
 
     write_info(dst_meta.info, dst_meta.root)
 

src/lerobot/datasets/io_utils.py

@@ -20,6 +20,7 @@ import datasets
 import numpy as np
 import pandas
 import pandas as pd
+import pyarrow as pa
 import pyarrow.dataset as pa_ds
 import pyarrow.parquet as pq
 import torch
@@ -270,21 +271,49 @@ def hf_transform_to_torch(items_dict: dict[str, list[Any]]) -> dict[str, list[to
     return items_dict
 
 
+def write_table_one_row_group_per_episode(table: pa.Table, path: Path) -> None:
+    """Write ``table`` with one parquet row group per episode (in episode order).
+
+    Keeps shards random-access friendly (``read_row_group(i)`` fetches episode i),
+    mirroring the recording writer. ``table`` must carry a contiguous
+    ``episode_index`` column.
+    """
+    episode_index = table.column("episode_index").to_numpy(zero_copy_only=False)
+    starts = np.concatenate(([0], np.nonzero(np.diff(episode_index))[0] + 1))
+    writer = pq.ParquetWriter(str(path), table.schema, compression="snappy", use_dictionary=True)
+    try:
+        for start, stop in zip(starts, np.append(starts[1:], len(episode_index)), strict=True):
+            writer.write_table(table.slice(start, stop - start))  # one episode -> one row group
+    finally:
+        writer.close()
+
+
 def to_parquet_with_hf_images(
     df: pandas.DataFrame, path: Path, features: datasets.Features | None = None
 ) -> None:
-    """This function correctly writes to parquet a panda DataFrame that contains images encoded by HF dataset.
-    This way, it can be loaded by HF dataset and correctly formatted images are returned.
+    """Write a DataFrame with HF-encoded images to parquet, one row group per episode.
 
-    Args:
-        df: DataFrame to write to parquet.
-        path: Path to write the parquet file.
-        features: Optional HuggingFace Features schema. If provided, ensures image columns
-                  are properly typed as Image() in the parquet schema.
+    Images are embedded into the arrow table first (``ParquetWriter.write_table``
+    does not embed external image files like ``Dataset.to_parquet`` does).
+    ``features`` types image columns as ``Image()`` in the parquet schema.
     """
-    # TODO(qlhoest): replace this weird synthax by `df.to_parquet(path)` only
     ds = datasets.Dataset.from_dict(df.to_dict(orient="list"), features=features)
-    ds.to_parquet(path)
+    ds = embed_images(ds)
+    table = ds.with_format("arrow")[:]
+    if "episode_index" in table.column_names:
+        write_table_one_row_group_per_episode(table, path)
+    else:
+        # No episode boundaries to align row groups to — keep a single write.
+        pq.write_table(table, str(path))
+
+
+def to_parquet_one_row_group_per_episode(df: pandas.DataFrame, path: Path) -> None:
+    """Write a (non-image) DataFrame to parquet with one row group per episode."""
+    table = pa.Table.from_pandas(df, preserve_index=False)
+    if "episode_index" in table.column_names:
+        write_table_one_row_group_per_episode(table, path)
+    else:
+        pq.write_table(table, str(path))
 
 
 def item_to_torch(item: dict) -> dict:

src/lerobot/scripts/lerobot_eval.py

@@ -72,8 +72,9 @@ from termcolor import colored
 from torch import Tensor, nn
 from tqdm import trange
 
-from lerobot.configs import parser
+from lerobot.configs import FeatureType, parser
 from lerobot.configs.eval import EvalPipelineConfig
+from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from lerobot.envs import (
     check_env_attributes_and_types,
     close_envs,
@@ -84,7 +85,7 @@ from lerobot.envs import (
 from lerobot.policies import PreTrainedPolicy, make_policy, make_pre_post_processors
 from lerobot.processor import PolicyProcessorPipeline
 from lerobot.types import PolicyAction
-from lerobot.utils.constants import ACTION, DONE, OBS_STR, REWARD
+from lerobot.utils.constants import ACTION, DONE, OBS_IMAGE, OBS_IMAGES, OBS_STR, REWARD
 from lerobot.utils.device_utils import get_safe_torch_device
 from lerobot.utils.import_utils import register_third_party_plugins
 from lerobot.utils.io_utils import write_video
@@ -95,6 +96,65 @@ from lerobot.utils.utils import (
 )
 
 
+def _env_features_to_dataset_features(env_features: dict) -> dict:
+    """Convert EnvConfig.features to the dict format expected by LeRobotDataset.create()."""
+    features = {}
+    for key, ft in env_features.items():
+        shape = tuple(ft.shape)
+        if ft.type is FeatureType.VISUAL:
+            features[key] = {"dtype": "video", "shape": shape, "names": ["height", "width", "channel"]}
+        else:
+            features[key] = {"dtype": "float32", "shape": shape, "names": None}
+    features["next.reward"] = {"dtype": "float32", "shape": (1,), "names": None}
+    features["next.success"] = {"dtype": "bool", "shape": (1,), "names": None}
+    features["next.done"] = {"dtype": "bool", "shape": (1,), "names": None}
+    return features
+
+
+def _build_raw_frame(
+    raw_obs: dict,
+    env_idx: int,
+    action: np.ndarray,
+    reward: float,
+    success: bool,
+    done: bool,
+    task: str,
+    env_features: dict,
+) -> dict:
+    """Build a dataset frame from raw env observations for one env index.
+
+    Keys in the frame match the keys in env_features so they align with the
+    dataset schema created by _env_features_to_dataset_features().
+    """
+    frame: dict[str, Any] = {}
+    for key in env_features:
+        if key == ACTION:
+            continue
+        if key.startswith("next."):
+            continue
+        if "pixels" in raw_obs and isinstance(raw_obs["pixels"], dict):
+            for cam_name, img in raw_obs["pixels"].items():
+                candidate = f"{OBS_IMAGES}.{cam_name}"
+                if candidate == key:
+                    frame[key] = img[env_idx]
+            if key in frame:
+                continue
+        if "pixels" in raw_obs and not isinstance(raw_obs["pixels"], dict) and key in ("pixels", OBS_IMAGE):
+            frame[key] = raw_obs["pixels"][env_idx]
+            continue
+        if key in raw_obs and isinstance(raw_obs[key], np.ndarray):
+            val = raw_obs[key][env_idx]
+            if val.dtype == np.float64:
+                val = val.astype(np.float32)
+            frame[key] = val
+    frame[ACTION] = action
+    frame["next.reward"] = np.atleast_1d(np.float32(reward))
+    frame["next.success"] = np.atleast_1d(np.bool_(success))
+    frame["next.done"] = np.atleast_1d(np.bool_(done))
+    frame["task"] = task
+    return frame
+
+
 def rollout(
     env: gym.vector.VectorEnv,
     policy: PreTrainedPolicy,
@@ -105,6 +165,10 @@ def rollout(
     seeds: list[int] | None = None,
     return_observations: bool = False,
     render_callback: Callable[[gym.vector.VectorEnv], None] | None = None,
+    recording_dir: Path | None = None,
+    env_features: dict | None = None,
+    recording_repo_id: str | None = None,
+    recording_private: bool = False,
 ) -> dict:
     """Run a batched policy rollout once through a batch of environments.
 
@@ -145,6 +209,33 @@ def rollout(
     if render_callback is not None:
         render_callback(env)
 
+    recording_datasets: list[LeRobotDataset] | None = None
+    raw_observation = None
+    task_desc = ""
+    if recording_dir is not None and env_features is not None:
+        features = _env_features_to_dataset_features(env_features)
+        fps = env.unwrapped.metadata.get("render_fps", 30)
+        recording_datasets = []
+        for i in range(env.num_envs):
+            multi_env = env.num_envs > 1
+            root = str(recording_dir / f"env_{i}") if multi_env else str(recording_dir)
+            base_repo_id = recording_repo_id or "eval_recording"
+            repo_id = f"{base_repo_id}_env_{i}" if multi_env else base_repo_id
+            recording_datasets.append(
+                LeRobotDataset.create(
+                    repo_id=repo_id,
+                    fps=fps,
+                    features=features,
+                    root=root,
+                    use_videos=True,
+                )
+            )
+        raw_observation = deepcopy(observation)
+        try:
+            task_desc = list(env.call("task_description"))[0]
+        except (AttributeError, NotImplementedError):
+            task_desc = ""
+
     all_observations = []
     all_actions = []
     all_rewards = []
@@ -217,6 +308,26 @@ def rollout(
         else:
             successes = [False] * env.num_envs
 
+        if recording_datasets is not None and raw_observation is not None:
+            prev_done = done.copy()
+            for env_idx in range(env.num_envs):
+                if prev_done[env_idx]:
+                    continue
+                frame = _build_raw_frame(
+                    raw_observation,
+                    env_idx,
+                    action_numpy[env_idx],
+                    reward[env_idx],
+                    successes[env_idx],
+                    bool(terminated[env_idx] | truncated[env_idx]),
+                    task_desc,
+                    recording_datasets[env_idx].features,
+                )
+                recording_datasets[env_idx].add_frame(frame)
+                if terminated[env_idx] or truncated[env_idx]:
+                    recording_datasets[env_idx].save_episode()
+            raw_observation = deepcopy(observation)
+
         # Keep track of which environments are done so far.
         # Mark the episode as done if we reach the maximum step limit.
         # This ensures that the rollout always terminates cleanly at `max_steps`,
@@ -255,6 +366,12 @@ def rollout(
             stacked_observations[key] = torch.stack([obs[key] for obs in all_observations], dim=1)
         ret[OBS_STR] = stacked_observations
 
+    if recording_datasets is not None:
+        for ds in recording_datasets:
+            ds.finalize()
+            if recording_repo_id is not None:
+                ds.push_to_hub(private=recording_private)
+
     if hasattr(policy, "use_original_modules"):
         policy.use_original_modules()
 
@@ -273,6 +390,10 @@ def eval_policy(
     videos_dir: Path | None = None,
     return_episode_data: bool = False,
     start_seed: int | None = None,
+    recording_dir: Path | None = None,
+    env_features: dict | None = None,
+    recording_repo_id: str | None = None,
+    recording_private: bool = False,
 ) -> dict:
     """
     Args:
@@ -361,6 +482,10 @@ def eval_policy(
             seeds=list(seeds) if seeds else None,
             return_observations=return_episode_data,
             render_callback=render_frame if max_episodes_rendered > 0 else None,
+            recording_dir=recording_dir,
+            env_features=env_features,
+            recording_repo_id=recording_repo_id,
+            recording_private=recording_private,
         )
 
         # Figure out where in each rollout sequence the first done condition was encountered (results after
@@ -563,6 +688,10 @@ def eval_main(cfg: EvalPipelineConfig):
     # Create environment-specific preprocessor and postprocessor (e.g., for LIBERO environments)
     env_preprocessor, env_postprocessor = make_env_pre_post_processors(env_cfg=cfg.env, policy_cfg=cfg.policy)
 
+    recording_dir = Path(cfg.output_dir) / "recordings" if cfg.eval.recording else None
+    max_episodes_rendered = 0 if cfg.eval.recording else 10
+    videos_dir = None if cfg.eval.recording else Path(cfg.output_dir) / "videos"
+
     with torch.no_grad(), torch.autocast(device_type=device.type) if cfg.policy.use_amp else nullcontext():
         info = eval_policy_all(
             envs=envs,
@@ -572,10 +701,15 @@ def eval_main(cfg: EvalPipelineConfig):
             preprocessor=preprocessor,
             postprocessor=postprocessor,
             n_episodes=cfg.eval.n_episodes,
-            max_episodes_rendered=10,
-            videos_dir=Path(cfg.output_dir) / "videos",
+            max_episodes_rendered=max_episodes_rendered,
+            videos_dir=videos_dir,
+            return_episode_data=False,
             start_seed=cfg.seed,
             max_parallel_tasks=cfg.env.max_parallel_tasks,
+            recording_dir=recording_dir,
+            env_features=cfg.env.features if cfg.eval.recording else None,
+            recording_repo_id=cfg.eval.recording_repo_id,
+            recording_private=cfg.eval.recording_private,
         )
         print("Overall Aggregated Metrics:")
         print(info["overall"])
@@ -618,6 +752,10 @@ def eval_one(
     videos_dir: Path | None,
     return_episode_data: bool,
     start_seed: int | None,
+    recording_dir: Path | None = None,
+    env_features: dict | None = None,
+    recording_repo_id: str | None = None,
+    recording_private: bool = False,
 ) -> TaskMetrics:
     """Evaluates one task_id of one suite using the provided vec env."""
 
@@ -635,6 +773,10 @@ def eval_one(
         videos_dir=task_videos_dir,
         return_episode_data=return_episode_data,
         start_seed=start_seed,
+        recording_dir=recording_dir,
+        env_features=env_features,
+        recording_repo_id=recording_repo_id,
+        recording_private=recording_private,
     )
 
     per_episode = task_result["per_episode"]
@@ -661,6 +803,10 @@ def run_one(
     videos_dir: Path | None,
     return_episode_data: bool,
     start_seed: int | None,
+    recording_dir: Path | None = None,
+    env_features: dict | None = None,
+    recording_repo_id: str | None = None,
+    recording_private: bool = False,
 ):
     """
     Run eval_one for a single (task_group, task_id, env).
@@ -672,7 +818,13 @@ def run_one(
         task_videos_dir = videos_dir / f"{task_group}_{task_id}"
         task_videos_dir.mkdir(parents=True, exist_ok=True)
 
-    # Call the existing eval_one (assumed to return TaskMetrics-like dict)
+    task_recording_dir = None
+    task_repo_id = None
+    if recording_dir is not None and env_features is not None:
+        task_recording_dir = recording_dir / f"{task_group}_{task_id}"
+        if recording_repo_id is not None:
+            task_repo_id = f"{recording_repo_id}_{task_group}_{task_id}"
+
     metrics = eval_one(
         env,
         policy=policy,
@@ -685,8 +837,12 @@ def run_one(
         videos_dir=task_videos_dir,
         return_episode_data=return_episode_data,
         start_seed=start_seed,
+        recording_dir=task_recording_dir,
+        env_features=env_features,
+        recording_repo_id=task_repo_id,
+        recording_private=recording_private,
     )
-    # ensure we always provide video_paths key to simplify accumulation
+
     if max_episodes_rendered > 0:
         metrics.setdefault("video_paths", [])
     return task_group, task_id, metrics
@@ -702,6 +858,10 @@ def eval_policy_all(
     n_episodes: int,
     *,
     max_episodes_rendered: int = 0,
+    recording_dir: Path | None = None,
+    env_features: dict | None = None,
+    recording_repo_id: str | None = None,
+    recording_private: bool = False,
     videos_dir: Path | None = None,
     return_episode_data: bool = False,
     start_seed: int | None = None,
@@ -761,6 +921,10 @@ def eval_policy_all(
         videos_dir=videos_dir,
         return_episode_data=return_episode_data,
         start_seed=start_seed,
+        recording_dir=recording_dir,
+        env_features=env_features,
+        recording_repo_id=recording_repo_id,
+        recording_private=recording_private,
     )
 
     if max_parallel_tasks <= 1:

tests/annotations/test_writer.py

@@ -28,6 +28,7 @@ import pytest
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
 
+import pandas as pd  # noqa: E402
 import pyarrow.parquet as pq  # noqa: E402
 
 from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
@@ -344,6 +345,78 @@ def test_annotation_metadata_sync_allows_non_streaming_load(
     assert len(dataset) == 24
 
 
+def _build_packed_dataset(root: Path, episode_lengths: list[int], *, fps: int = 10) -> Path:
+    """Pack several episodes into a single shard (vs build_annotation_dataset's one-per-file),
+    so the writer's rewrite must re-emit one row group per episode instead of collapsing them."""
+    from lerobot.datasets.io_utils import write_tasks
+    from lerobot.utils.io_utils import write_json
+
+    data_dir = root / "data" / "chunk-000"
+    data_dir.mkdir(parents=True, exist_ok=True)
+
+    episode_index, frame_index, timestamp, task_index, subtask_index = [], [], [], [], []
+    for ep, length in enumerate(episode_lengths):
+        episode_index += [ep] * length
+        frame_index += list(range(length))
+        timestamp += [round(i / fps, 6) for i in range(length)]
+        task_index += [0] * length
+        subtask_index += [0] * length  # legacy column the writer must drop
+    pd.DataFrame(
+        {
+            "episode_index": episode_index,
+            "frame_index": frame_index,
+            "timestamp": timestamp,
+            "task_index": task_index,
+            "subtask_index": subtask_index,
+        }
+    ).to_parquet(data_dir / "file-000.parquet", index=False)
+
+    tasks_df = pd.DataFrame({"task_index": [0]}, index=pd.Index(["do the thing"], name="task"))
+    write_tasks(tasks_df, root)
+    write_json(
+        {"codebase_version": "v3.1", "fps": fps, "features": {}, "total_episodes": len(episode_lengths)},
+        root / "meta" / "info.json",
+    )
+    return root
+
+
+def test_writer_one_row_group_per_episode(tmp_path: Path) -> None:
+    """Rewriting a packed shard must keep one row group per episode, not collapse
+    every episode into a single giant row group."""
+    episode_lengths = [4, 6, 5]  # unequal lengths, all in one shard
+    root = _build_packed_dataset(tmp_path / "ds", episode_lengths)
+    shard = root / "data" / "chunk-000" / "file-000.parquet"
+    assert pq.ParquetFile(shard).metadata.num_row_groups == 1, "fixture should start collapsed"
+
+    staging_dir = tmp_path / "stage"
+    for ep in range(len(episode_lengths)):
+        _stage_episode(
+            staging_dir,
+            ep,
+            plan=[
+                {
+                    "role": "assistant",
+                    "content": f"subtask for ep {ep}",
+                    "style": "subtask",
+                    "timestamp": 0.0,
+                    "tool_calls": None,
+                }
+            ],
+        )
+
+    records = list(iter_episodes(root))
+    LanguageColumnsWriter().write_all(records, staging_dir, root)
+
+    # One row group per episode, with row counts matching the episode lengths.
+    md = pq.ParquetFile(shard).metadata
+    assert md.num_row_groups == len(episode_lengths)
+    assert [md.row_group(i).num_rows for i in range(md.num_row_groups)] == episode_lengths
+    # Language columns are still present after the per-episode rewrite.
+    table = pq.read_table(shard)
+    assert "language_persistent" in table.column_names
+    assert "language_events" in table.column_names
+
+
 def test_speech_atom_shape_matches_plan_spec() -> None:
     atom = speech_atom(2.5, "I'm cleaning up!")
     assert atom["role"] == "assistant"

tests/datasets/test_aggregate.py

@@ -32,6 +32,26 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from tests.fixtures.constants import DUMMY_REPO_ID
 
 
+def assert_data_shards_one_row_group_per_episode(root):
+    """Every aggregated DATA shard must have exactly one parquet row group per episode."""
+    import pyarrow.parquet as pq
+
+    shards = sorted((root / "data").rglob("*.parquet"))
+    assert shards, f"no data shards found under {root}/data"
+    n_episodes = 0
+    for shard in shards:
+        pf = pq.ParquetFile(shard)
+        episodes = pf.read(columns=["episode_index"]).column("episode_index").to_pylist()
+        assert pf.metadata.num_row_groups == len(set(episodes)), shard
+        for i in range(pf.metadata.num_row_groups):
+            rg_episodes = set(
+                pf.read_row_group(i, columns=["episode_index"]).column("episode_index").to_pylist()
+            )
+            assert len(rg_episodes) == 1, f"{shard} row group {i} spans episodes {rg_episodes}"
+        n_episodes += len(set(episodes))
+    return n_episodes
+
+
 def assert_episode_and_frame_counts(aggr_ds, expected_episodes, expected_frames):
     """Test that total number of episodes and frames are correctly aggregated."""
     assert aggr_ds.num_episodes == expected_episodes, (
@@ -566,6 +586,41 @@ def assert_image_frames_integrity(aggr_ds, ds_0, ds_1):
             )
 
 
+@pytest.mark.parametrize("use_videos", [True, False], ids=["video", "image"])
+def test_aggregate_one_row_group_per_episode(tmp_path, lerobot_dataset_factory, use_videos):
+    """Aggregated DATA shards keep one row group per episode (not one collapsed group).
+
+    Covers both the non-image (``df.to_parquet``) and image
+    (``to_parquet_with_hf_images``) write branches, including the merge-into-
+    existing-file branch via a low file-size threshold that forces packing.
+    """
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "rg_0",
+        repo_id=f"{DUMMY_REPO_ID}_rg_0",
+        total_episodes=3,
+        total_frames=60,
+        use_videos=use_videos,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "rg_1",
+        repo_id=f"{DUMMY_REPO_ID}_rg_1",
+        total_episodes=4,
+        total_frames=80,
+        use_videos=use_videos,
+    )
+
+    aggr_root = tmp_path / "rg_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_rg_aggr",
+        aggr_root=aggr_root,
+    )
+
+    n_episodes = assert_data_shards_one_row_group_per_episode(aggr_root)
+    assert n_episodes == ds_0.num_episodes + ds_1.num_episodes
+
+
 def test_aggregate_image_datasets(tmp_path, lerobot_dataset_factory):
     """Test aggregation of image-based datasets preserves HuggingFace Image schema.
 

tests/datasets/test_datasets.py

@@ -51,7 +51,7 @@ from lerobot.robots import make_robot_from_config
 from lerobot.transforms import ImageTransforms, ImageTransformsConfig
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, OBS_STR, REWARD
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_REPO_ID
+from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_MOTOR_FEATURES, DUMMY_REPO_ID
 from tests.mocks.mock_robot import MockRobotConfig
 from tests.utils import require_x86_64_kernel
 
@@ -133,6 +133,21 @@ def test_dataset_feature_with_forward_slash_raises_error():
         )
 
 
+def test_create_does_not_mutate_input_features(tmp_path, empty_lerobot_dataset_factory):
+    # ``create`` must deep-copy features so a dataset built from another's features stays independent.
+    dataset = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds1", features=DUMMY_MOTOR_FEATURES, use_videos=False
+    )
+    dataset_copy = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds2", features=dataset.meta.features, use_videos=False
+    )
+
+    original_shape = dataset.meta.info.features["state"]["shape"]
+    dataset_copy.meta.info.features["state"]["shape"] = (999,)
+
+    assert dataset.meta.info.features["state"]["shape"] == original_shape
+
+
 def test_add_frame_missing_task(tmp_path, empty_lerobot_dataset_factory):
     features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
     dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)