Skip to content

README.md

Latest commit

 

History

History
90 lines (60 loc) · 2.42 KB

README.md

File metadata and controls

90 lines (60 loc) · 2.42 KB

lmdb2lerobot

Convert simulation Step-4 output (lmdb/ + meta_info.pkl) into a LeRobot v2 dataset.

One-shot workflow

run_local.sh runs three stages by default:

  1. LMDB → LeRobot (lmdb2lerobot_arx_sim.py)
  2. sim2real parquet post-process (sim2real.py) — skip with --skip-sim2real
  3. Remove near-static frames (remove_static_frames.py, in-place on the dataset) — skip with --keep-static-frames

1. Environment (once)

bash components/lmdb2lerobot/setup_conda_lerobot.sh
conda activate lerobot

First run downloads PyTorch and LeRobot; expect on the order of 10–30 minutes.

2. Convert one directory

From the repository root (edit paths to your session):

conda activate lerobot
bash components/lmdb2lerobot/run_local.sh \
  --src ./replay/<session>/out_updated \
  --out ./replay/<session>/lerobot_dataset
  • --src: full out_updated tree, or a single episode folder that contains lmdb/.
  • --out: LeRobot dataset root; removed and recreated if it already exists.
  • Static frames: removed automatically after step 2 (tune with --static-threshold-ratio, --static-workers; use --keep-static-frames to keep all frames).

More flags: bash components/lmdb2lerobot/run_local.sh --help

Manual — remove_static_frames only

If you already have a LeRobot dataset and want to trim static frames without re-running the full script (or to re-run with different thresholds):

conda activate lerobot
python components/lmdb2lerobot/remove_static_frames.py /path/to/lerobot_dataset

Back up the dataset first — this edits parquet, videos, and meta/ in place.

Optional — batch conversion

conda activate lerobot
bash components/lmdb2lerobot/run_batch.sh --scan /path/to/parent --out-root /path/to/lerobot_outputs --workers 4

Details: bash components/lmdb2lerobot/run_batch.sh --help

Input layouts (auto-detected)

out_updated/
  000000/lmdb/data.mdb
  000001/lmdb/data.mdb
  ...

# or single-episode folder with lmdb/
# or legacy .../out/<timestamp>/lmdb/data.mdb

Output layout (summary)

lerobot_dataset/
  data/chunk-000/episode_*.parquet
  videos/chunk-000/images.rgb.{head,hand_left,hand_right}/episode_*.mp4
  meta/info.json, episodes.jsonl, tasks.jsonl

observation.state and action are 14-D: left arm (×6) + left gripper (×1) + right arm (×6) + right gripper (×1).