Datasets:

---

lerobot

/

droid_1.0.1

like 19

Follow

LeRobot 4.48k

Title:

Hi, I found a metadata/data inconsistency in lerobot/droid_1.0.1 that can cause LeRobotDataset(..., episodes=...) to load incorrect extra samples.

Dataset revision:

bd92a2c4d6c85f9e16bd7c1f9729f5496d8be672

Problem

The dataset repository contains 156 data parquet shards under:

data/chunk-000/file-000.parquet
...
data/chunk-000/file-155.parquet

However, meta/episodes only references 86 of them through the episode-level columns:

data/chunk_index
data/file_index

Reading all meta/episodes/chunk-000/*.parquet shows:

unique data/chunk_index: [0]
unique data/file_index count: 86
data/file_index min/max: 0 / 85
contains file_index 86: False
contains file_index 155: False

So according to meta/episodes, only these shards are part of the dataset:

data/chunk-000/file-000.parquet
...
data/chunk-000/file-085.parquet

But the repository also contains:

data/chunk-000/file-086.parquet
...
data/chunk-000/file-155.parquet

These extra shards are not referenced by meta/episodes.

Evidence

The metadata is internally consistent:

meta.total_episodes = 95,658
meta.total_frames = 27,630,375
sum(meta/episodes.length) = 27,630,375

But the physical data parquet files contain more rows:

number of data parquet files on disk: 156
sum(rows in all data/*.parquet): 39,995,463

The extra parquet files are not empty. For example:

data/chunk-000/file-155.parquet
rows: 39,695
episode_index: 95477..95616
index: 27578956..27618650

But these episodes are already described in meta/episodes as belonging to data/file_index = 85, with different index ranges and lengths.

Example:

episode 95477 in file-155.parquet:
rows: 131
index range: 27578956..27579086
frame_index: 0..130

episode 95477 in meta/episodes:
length: 166
dataset_from_index: 27580071
dataset_to_index: 27580237
data/file_index: 85

Another example:

episode 95616 in file-155.parquet:
rows: 255
index range: 27618396..27618650

episode 95616 in meta/episodes:
length: 130
dataset_from_index: 27618177
dataset_to_index: 27618307
data/file_index: 85

This suggests that file-086.parquet to file-155.parquet are stale, duplicate, or otherwise unreferenced shards that should not be loaded for the current metadata.

Loader issue

lerobot.datasets.io_utils.load_nested_dataset() currently loads parquet files by globbing all files under data/:

paths = sorted(pq_dir.glob("/.parquet"))
filters = pa_ds.field("episode_index").isin(episodes) if episodes is not None else None
return Dataset.from_parquet([str(path) for path in paths], filters=filters, features=features)

This means that when episodes is provided, the filter is applied to all 156 parquet files, including the 70 unreferenced shards.

As a result, LeRobotDataset(..., episodes=success_episode_ids) loads rows from unreferenced shards if their episode_index matches the requested list.

For example, selecting successful episodes from meta/episodes gives:

successful episodes: 78,864
sum(length for successful episodes in meta/episodes): 23,408,858

But loading with episodes=successful_episode_ids returns:

filtered dataset length: 33,569,891
unique episodes: 78,864

This is larger than the metadata-defined successful subset and even larger than expected from meta/episodes.

Impact

This can silently corrupt episode-filtered datasets. In my case, training with success_only=True loaded extra rows from unreferenced shards and caused a CUDA device-side assert during the first forward pass. Training with success_only=False did not
trigger the issue because the dataset length follows meta.total_frames.

The root cause appears to be that episode filtering reads all parquet files in data/, instead of only the parquet files referenced by meta/episodes.

Expected behavior

LeRobotDataset should only load data parquet shards that are referenced by meta/episodes, or the dataset repository should not contain unreferenced data shards.

For this revision, the loader should only consider:

data/chunk-000/file-000.parquet
...
data/chunk-000/file-085.parquet

Suggested fixes

Possible fixes:

1. Remove unreferenced data shards from lerobot/droid_1.0.1 for this revision, if they are stale.
2. Update load_nested_dataset() / DatasetReader to restrict parquet paths to those referenced by meta/episodes.
3. Add a dataset consistency check that verifies:
   • all meta/episodes referenced data shards exist
   • no unreferenced data shards are loaded
   • sum(rows of loaded data shards) matches meta.total_frames

Thanks.