LeWAM/lewam-drawer-3view

LeWAM Drawer Cleanup — three views

Simulation training data used by LeWAM.

• Episodes: 1,026
• Frames: 298,235
• Camera columns: pixels, pixels_r0eih, pixels_r1eih
• Action dimensions: 22

Files

File	Bytes	Purpose
drawer_fixed.h5	18,340,768,820	Scene camera, actions and episode metadata
drawer_3view.h5	42,575,765,750	Wrist-camera supplement
drawer_multiview.h5	11,073,170	Multiview entry point and model XML; requires both companion HDF5 files

Download the entire repository and keep the HDF5 files together. The multiview entry point uses relative HDF5 links to its companion files; it contains no cluster-specific paths. Open drawer_multiview.h5 for all views.

Download and read

pip install huggingface_hub h5py hdf5plugin

from pathlib import Path
from huggingface_hub import snapshot_download
import hdf5plugin  # registers the image compression filters
import h5py

root = Path(snapshot_download("LeWAM/lewam-drawer-3view", repo_type="dataset"))
with h5py.File(root / "drawer_multiview.h5", "r") as data:
    print(list(data.keys()))
    image = data["pixels"][0]
    action = data["action"][0]

Integrity and format

manifest.json records the byte size and SHA-256 of every HDF5 file. Camera frames are 224 × 224 RGB. ep_offset and ep_len describe episode boundaries. No optional preload cache is required.

model_xml contains 1,026 MuJoCo XML strings, one per episode, indexed by ep_idx. The strings come from the single-view Drawer dataset and restore each trajectory's payload geometry for evaluation. drawer_multiview.h5 stores this column directly with lossless gzip compression. Its camera and trajectory columns use the existing relative links to the two companion files.

Camera frames, actions, states and episode boundaries retain the source training data.