Aetherock

A task-organized vision–tactile manipulation dataset with synchronized stored wrist-camera observations, two tactile image channels, two force channels, wrist poses, gripper values, and task instructions. The original files use the FTP1-compatible Zarr v2 layout for contribution to Open-X-Tactile.

Browse real samples in Data Studio by task, episode ID, or frame index. The original Zarr archives retain the complete recordings.

Browse real frames in Data Studio · Full episode index · Task catalog · Original files

Full source dataset: 11 archives · 3,760 stored episodes · 736,625 timesteps · 138.19 GB.

Data Studio inspection layer: 264 sampled frames from 33 episodes. The image table is a preview; it is not the complete training dataset. The episode index covers every stored episode.

Explore the data

Data Studio subset	Rows	What you can inspect
Real frame preview	264	Images, task descriptions, episode/frame IDs and original numeric signals
Complete episode index	3,760	Every episode's source archive, start/end frame, length and available FPS
Task / archive catalog	11	A representative observation and full counts for each original archive

Use the subset selector in Data Studio, then filter by task, collection_mode, or episode_id. Image cells contain actual stored pixels. Position, gripper and force columns use the corresponding original frame values. Source archive paths and SHA-256 hashes are included for traceability.

Episode observations

Eight evenly spaced observations from selected original episodes. Animation speed is illustrative, not the original recording rate. Each tile combines a wrist RGB image and the two tactile channels from the same stored timestep.

Full-dataset coverage

Counts are computed from the original arrays, independently of the small frame preview. The erase_board and eraseboard directories remain separate source entries.

An episode's numeric signals

The plot uses eight real observations from one episode. Lines connect sampled values for visualization; coordinates and signals are shown in their stored conventions.

Observation gallery across every original archive

Original archives

File	Collection	Episodes	Timesteps	Size (GB)
bread_moving.zarr.zip	tactile manipulation	199	20,597	3.73
clamp_seal.zarr.zip	tactile manipulation	229	34,630	6.98
close_bottle.zarr.zip	tactile manipulation	497	141,017	26.94
erase_board.zarr.zip	tactile manipulation	498	95,345	17.58
eraseboard.zarr.zip	tactile manipulation	215	31,397	5.58
peg_flowers.zarr.zip	tactile manipulation	212	29,663	5.93
screw.zarr.zip	tactile manipulation	501	95,080	17.48
stiring.zarr.zip	tactile manipulation	497	88,894	16.35
stowing.zarr.zip	tactile manipulation	496	149,674	28.13
towel_hanging.zarr.zip	tactile manipulation	206	33,125	6.24
yellow_to_pink.zarr.zip	tactile manipulation	210	17,203	3.24
Total		3,760	736,625	138.19

Sizes are rounded decimal GB. Exact sizes, source hashes, array schemas and sample indices are recorded in showcase_manifest.json. Statistics were verified from the original archives on 2026-10-27.

Modalities and fields

Modality	Original array	Data Studio column
Wrist RGB	data/right_wrist_camera_rgb	image
Source left/right tactile RGB	data/right_tactile_data_gripper[:, 0/1]	tactile_left, tactile_right
Source left/right force	data/right_tactile_data_force	force_left, force_right
Wrist position + rotation vector	data/right_wrist_pose	position_x/y/z, orientation_axis_angle
Hand/gripper scalar	data/right_hand_joints	gripper_value
Task instruction	data/sub_task_instruction	description, source_instruction
Source timestamp	data/timestamps	source_timestamp
Episode boundaries	meta/episode_ends	Full episode_index table

The tactile/force channel order is read from the archive attributes (source left, source right). Numeric values retain the source conventions. The preview does not define new coordinate frames or force units.

Download and load

This repository uses automatic access requests. Request access on the dataset page, then authenticate with hf auth login before downloading or querying the previews.

pip install -U huggingface_hub datasets "zarr<3"

Load the small inspection tables

from datasets import load_dataset

frames = load_dataset("lihong-cs/aetherock", "frame_preview", split="preview", streaming=True, token=True)
sample = next(iter(frames))
print(sample["task"], sample["episode_id"], sample["frame_index"])
sample["image"].save("preview-rgb.png")

episodes = load_dataset("lihong-cs/aetherock", "episode_index", split="episodes", streaming=True, token=True)
print(next(iter(episodes)))

Read an original Zarr archive

Download one complete source archive:

hf download lihong-cs/aetherock bread_moving/bread_moving.zarr.zip --repo-type dataset --local-dir ./data

import zarr

path = "./data/bread_moving/bread_moving.zarr.zip"
with zarr.ZipStore(path, mode="r") as store:
    root = zarr.open_group(store=store, mode="r", path="bread_moving.zarr")
    episode_ends = root["meta/episode_ends"][:]
    episode_index = 0
    start = 0 if episode_index == 0 else int(episode_ends[episode_index - 1])
    stop = int(episode_ends[episode_index])
    wrist_pose = root["data/right_wrist_pose"][start:stop]
    force = root["data/right_tactile_data_force"][start:stop]
    rgb = root["data/right_wrist_camera_rgb"][start]
    tactile_pair = root["data/right_tactile_data_gripper"][start]
    print(wrist_pose.shape, force.shape, rgb.shape, tactile_pair.shape)

Read image arrays by individual frame or short slice. Downloading an archive retrieves the full original file; loading frame_preview retrieves the small inspection subset. Use episode-level splits for training and evaluation.

Preview construction and provenance

• The complete episode index comes from every meta/episode_ends array. Counts and array lengths are checked for consistency.
• Frame previews sample the first, middle and final stored episode in each archive, with eight evenly spaced timesteps per selected episode.
• Each preview image is stored losslessly in Parquet with a typed Hugging Face Image feature. Frame indices and source hashes identify the corresponding original data.
• Preview images and plots are additional inspection assets. Original Zarr archive contents are unchanged.
• time_in_episode_s_nominal and duration_s_nominal are derived only where FPS metadata exists. They are not replacement source timestamps.

Original source: ModelScope: lihongcs/aetherock.

License

This dataset is licensed under CC BY 4.0. Commercial use is permitted under the license terms, including appropriate attribution, a link to the license, and an indication of any changes.