SeonghuJeon/libero-gt-depth-aligned-hide-sites-fast

LIBERO GT Depth Aligned Hide Sites Fast

This dataset contains ground-truth depth sidecar files for LIBERO demonstrations.

Each episode is stored as a compressed .npz file. The depth is aligned to the original LIBERO HDF5 frames by re-rendering each frame after restoring the simulator state for that timestep.

Contents

• —episode_*.npz: per-episode depth and camera geometry
• —manifest_shard_*.jsonl: per-episode metadata, including the source HDF5 path
• —summary_shard_*.json: shard-level summary and generation settings

NPZ Keys

• —depth_raw: raw depth buffer values, shape (T, 2, 256, 256)
• —depth_meters: metric depth in meters, shape (T, 2, 256, 256)
• —camera_intrinsics: camera intrinsics, shape (T, 2, 3, 3)
• —camera_extrinsics_c2w: camera-to-world extrinsics, shape (T, 2, 4, 4)
• —frame_indices: frame indices within the source demo
• —camera_names: camera names, usually agentview and robot0_eye_in_hand
• —episode_index: global episode index

Cameras

• —agentview
• —robot0_eye_in_hand

Notes

• —The source LIBERO HDF5 files contain RGB observations and simulator states, but not these depth sidecars.
• —The manifest_shard_*.jsonl files record the source HDF5 path for every episode.
• —The camera geometry follows robosuite conventions.

Example

from huggingface_hub import snapshot_download
import numpy as np
import os

root = snapshot_download(
    repo_id="SeonghuJeon/libero-gt-depth-aligned-hide-sites-fast",
    repo_type="dataset",
)

sample_name = "episode_000000__KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it_demo__demo_0.npz"
sample = np.load(os.path.join(root, sample_name))

print(sample.files)
print(sample["depth_meters"].shape)
print(sample["camera_intrinsics"].shape)
print(sample["camera_extrinsics_c2w"].shape)