SeonghuJeon/droid-rgb-loader

DROID RGB Dataloader (Standalone)

Minimal, self-contained PyTorch loader for DROID video (RGB only). Pairs with our preprocessed DROID mirror:

Dataset: SeonghuJeon/droid-1.0.1-preprocessed

DO NOT use lerobot/droid_1.0.1 directly. That HF release has:

1. Broken from_timestamp / to_timestamp fields in meta/episodes/*.parquet — UNIX-epoch drift up to ~54 years. Any naive video seek lands on garbage.
2. 30,815 episodes (32%) with missing/truncated mp4 files that throw random av.error.* deep into training.

Our preprocessed mirror bakes in the timestamp fix and ships our blacklist + non-idle range sidecars. Same LeRobot v3.0 layout, drop-in replacement.

What this repo (droid-rgb-loader) ships

File	Purpose
droid_rgb_dataset.py	DroidRGBDataset, ~275 LoC, only deps are torch/numpy/pandas/pyarrow/pyav
example_usage.py	End-to-end sanity check + DataLoader example
_stats/droid_blacklist_eps.json	30,815 episodes to skip
_stats/droid_nonidle_ranges.json	per-episode {start,end} frame range where arm is moving

The stats files are also mirrored inside droid-1.0.1-preprocessed for convenience, but you can feed either copy to the loader via stats_dir=.

Step 1 — Download the preprocessed DROID dataset

HF_XET_HIGH_PERFORMANCE=1 \
hf download SeonghuJeon/droid-1.0.1-preprocessed \
  --repo-type dataset \
  --local-dir /path/to/droid-1.0.1-preprocessed

This is ~1.2TB (all three cameras + framecache sidecars + fixed meta). Video-only download (skip framecache, keep raw mp4 + meta) is smaller:

hf download SeonghuJeon/droid-1.0.1-preprocessed \
  --repo-type dataset \
  --local-dir /path/to/droid-1.0.1-preprocessed \
  --include "meta/*" "data/*" "videos/**/*.mp4"

If you only want one camera:

hf download SeonghuJeon/droid-1.0.1-preprocessed \
  --repo-type dataset \
  --local-dir /path/to/droid-1.0.1-preprocessed \
  --include "meta/*" "data/*" "videos/observation.images.exterior_2_left/**/*.mp4"

Step 2 — Install dependencies

pip install torch numpy pandas pyarrow av pillow

av is PyAV, the Python binding for ffmpeg. That's the only decode dep — no decord, no torchcodec, no torchvision VideoReader.

Step 3 — Get this loader repo

hf download SeonghuJeon/droid-rgb-loader --repo-type dataset --local-dir ./droid-rgb-loader
cd droid-rgb-loader

Step 4 — Run the example

python example_usage.py \
  --root /path/to/droid-1.0.1-preprocessed \
  --n-frames 8 --stride 3 --num-workers 2

Expected output:

usable episodes: 63908
exterior: shape=(2, 8, 224, 224, 3) dtype=torch.uint8
wrist:    shape=(2, 8, 224, 224, 3) dtype=torch.uint8

If usable episodes is much lower than ~64k, something is wrong with paths or stats dir.

Using DroidRGBDataset in your own code

from droid_rgb_dataset import DroidRGBDataset
from torch.utils.data import DataLoader

ds = DroidRGBDataset(
    root="/path/to/droid-1.0.1-preprocessed",
    stats_dir="./_stats",
    camera_keys=("observation.images.exterior_2_left",
                 "observation.images.wrist_left"),
    n_frames=16,          # window length
    stride=1,             # 1=15Hz, 3=5Hz
    image_size=(224, 224),
    apply_blacklist=True, # mandatory
    apply_nonidle=True,   # skip idle pre/post-roll
)

loader = DataLoader(ds, batch_size=4, num_workers=4, shuffle=True)
for batch in loader:
    ext = batch["observation.images.exterior_2_left"]  # (B, T, H, W, 3) uint8
    ...

Each __getitem__ picks a random valid window from the i-th usable episode. __len__ returns the number of usable episodes. Iterate the DataLoader multiple times per epoch if you want more diverse windows than one per episode.

Why the blacklist is non-negotiable

DROID on the original HF release has ~30,815 episodes (32%) whose mp4 files are missing, truncated, or fail to decode. They remain silently present in the metadata parquets. We rebuilt the blacklist by walking every observation.images.* mp4 on disk and we re-encode nothing — we just skip those episodes. The preprocessed mirror on SeonghuJeon/droid-1.0.1-preprocessed only contains video files we verified decode end-to-end, but the blacklist is still applied at dataloader time as a belt-and-suspenders check.

Why apply_nonidle matters

DROID was collected via VR teleop, and episodes start/end with the operator sitting still (several seconds of no-op frames). droid_nonidle_ranges.json tells you the [start, end) frame range where the arm is actually moving, so you don't waste training capacity on motionless pre-roll. We built it from gripper + cartesian velocity thresholds.

88,097 episodes (of the non-blacklisted ones) have a non-empty non-idle range. Episodes missing from this file are considered fully idle and dropped by the loader.

Camera keys

DROID stores three exterior views. We strongly recommend:

• observation.images.exterior_2_left (primary third-person view)
• observation.images.wrist_left (wrist-mounted)

Do NOT use observation.images.exterior_1_left: it is mis-calibrated or occluded on a large fraction of episodes. We blocked it in all our runs.

File layout (LeRobot v3.0)

<root>/
  meta/info.json
  meta/episodes/chunk-NNN/file-NNN.parquet    <- FIXED timestamps (mirror only)
  data/chunk-NNN/file-NNN.parquet
  videos/<camera_key>/chunk-NNN/file-NNN.mp4
  videos/<camera_key>/chunk-NNN/file-NNN.framecache/  (optional speedup, mirror only)

One mp4 holds many episodes. Each episode's window inside the mp4 is given by videos/<key>/from_timestamp in the episode meta parquet. The loader seeks there and decodes n_frames frames at the requested stride.

Performance notes

• PyAV seek + decode: ~10-30ms per window on a local SSD. Fast enough for training at num_workers=8+.
• Stride 3 = 5Hz matches the DROID effective control rate most policies use (native 15Hz is over-sampled for slow VR teleop data).
• Framecache: the preprocessed mirror ships .framecache sidecars next to each mp4 (one JPEG per frame, pre-decoded). This loader does not use them, but if you want roughly 3× faster random-window reads, ask us for the framecache reader — it's not included here to keep the code standalone.
• Multi-worker: PyAV containers are not fork-safe, so the loader opens and closes a container per __getitem__. This is the safe default.

Known limitations

• No language instructions returned (DROID has them in the data parquet; add a column read in _build_index if you need them).
• No action / state returned. If you need them later, the spec lives in meta/info.json → features, and the relevant columns are in data/chunk-NNN/file-NNN.parquet keyed by (episode_index, frame_index).
• Windows are per-episode random at __getitem__ time (not exhaustively enumerated). len(ds) == num_usable_episodes, not num_windows.

Contact

Built by Seonghu Jeon. If you hit a broken mp4 that's not in the blacklist, or a non-idle range that looks wrong, please send the episode index back so we can update the stats.