SeonghuJeon/droid-1.0.1-preprocessed

DROID 1.0.1 — Preprocessed Mirror

A ready-to-use mirror of lerobot/droid_1.0.1 with two preprocessing fixes baked in:

1. Fixed meta/episodes/*.parquet timestamps. The upstream release stores videos/*/from_timestamp as UNIX-epoch-drifted floats — up to ~54 years off. Any naive av.Container.seek() lands on garbage. We recomputed every from_timestamp / to_timestamp from actual video durations, backed up the original, and shipped the fixed parquets in this mirror.
2. Broken-episode blacklist applied at source. On the upstream release, 30,815 episodes (32%) have mp4 files that are missing, truncated, or fail to decode. This mirror contains only videos that pass a full PyAV decode, and the upstream episode indices are preserved (so the JSON sidecar below still matches).

Layout

Identical to LeRobot v3.0, drop-in compatible with any LeRobot v3 loader that reads meta/episodes/, data/, and videos/:

droid-1.0.1-preprocessed/
  meta/
    info.json
    tasks.parquet
    stats.json
    episodes/chunk-NNN/file-NNN.parquet      # FIXED timestamps
  data/chunk-NNN/file-NNN.parquet            # frame-level action/state (unchanged)
  videos/
    observation.images.exterior_2_left/chunk-NNN/file-NNN.mp4
    observation.images.exterior_2_left/chunk-NNN/file-NNN.framecache/   # optional
    observation.images.wrist_left/...
    observation.images.exterior_1_left/...
  _stats/
    droid_blacklist_eps.json                  # 30,815 broken episodes
    droid_nonidle_ranges.json                 # per-episode motion range

Framecache sidecars (.framecache/ directories next to each .mp4) are optional JPEG-per-frame caches we built during training to skip PyAV decode. You can ignore them if you just want raw mp4 decoding.

How to download

Everything (~1.2 TB):

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

Video-only, skip framecache (substantially 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" "_stats/*"

Only one camera (e.g. third-person exterior_2_left):

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" \
           "_stats/*"

Companion loader

For RGB-only PyTorch training, use the standalone loader we ship at SeonghuJeon/droid-rgb-loader. It expects this mirror as the root and reads the blacklist / non-idle sidecars automatically.

Camera guidance

• observation.images.exterior_2_left — primary third-person view, use this.
• observation.images.wrist_left — wrist-mounted, use this.
• observation.images.exterior_1_left — mis-calibrated or occluded on a large fraction of episodes. Blocked in all our runs.

Citation

DROID upstream: https://droid-dataset.github.io/

If this mirror saves you time, a ⭐ on the repo page is appreciated.