Datasets:

tqliu
/

gripper_data_share

license: cc-by-4.0
task_categories:
  - robotics
tags:
  - robotics
  - manipulation
  - rh20t
  - droid
  - agibot
pretty_name: Gripper Data Share (RH20T + DROID complete episodes + AgiBot usable lists)

Gripper Data Share

Curated robot-manipulation data for sharing. Two parts:

Complete episodes of RH20T and DROID, packed as .tar shards — every episode included here has all camera-view videos + a text caption + the action (robot state) data. Any episode missing one of these three was dropped.
Usable-episode lists for AgiBotWorld-Beta and AgiBotWorld2026 (these two datasets are not shipped here — only the lists of episodes whose FK skeleton projected correctly, so you can filter the original datasets).

File(s)	What
`RH20T_000.tar` … `RH20T_005.tar`	RH20T complete episodes (6 shards)
`droid_000.tar` … `droid_033.tar`	DROID complete episodes (34 shards)
`meta_000.tar`	Shared assets: RH20T `calib/` + `urdf/` + `task_description.json`, DROID `aggregated-annotations` + `manifest.json`
`AgiBotWorld-Beta_usable_episodes.txt`	146,771 usable AgiBotWorld-Beta episodes (`task/episode`)
`AgiBotWorld2026_usable_episodes.txt`	19,329 usable AgiBotWorld2026 episodes (`Domain/Scene/task/range/ep`)

Each shard is ≤ 25 GB; total tar data ≈ 0.94 TiB.

Download & extract

huggingface-cli download tqliu/gripper_data_share --repo-type dataset --local-dir gripper_data_share
cd gripper_data_share
for f in *.tar; do tar -xf "$f"; done      # reconstructs RH20T/ and droid/

Extraction yields:

RH20T/RH20T_cfg{5,6,7}/<episode>/...        # + shared calib/, urdf/, task_description.json
droid/<lab>/<success|failure>/<date>/<episode>/...   # + aggregated-annotations, manifest.json

Episode contents

RH20T — RH20T/RH20T_cfg{5,6,7}/task_..._cfg_N/

cam_<serial>/color.mp4 + cam_<serial>/timestamps.npy — all camera views (640×360)
transformed/ — action & state: joint.npy, tcp.npy, tcp_base.npy, gripper.npy, force_torque*.npy, high_freq_data.npy (+ tactile.npy where present)
metadata.json — calibration id, camera serials, timestamps
caption.json — {task_id, user_id, scene_id, cfg, task_description_english, task_description_chinese}. RH20T captions are task-level (all episodes of a task share the sentence), joined from task_description.json by the task id in the folder name (100% coverage).
Camera intrinsics/extrinsics are in the shared RH20T/RH20T_cfg{N}/calib/<calib_id>/ (look up metadata.json["calib"]).

DROID — droid/<lab>/<success|failure>/<date>/<episode>/

recordings/MP4/<serial>.mp4 — 3 camera views (2 exterior + 1 wrist, 1280×720)
trajectory.h5 — action & state (joint positions, cartesian, gripper, timestamps)
metadata_<uuid>.json
caption.json — {uuid, language_instructions}. DROID captions are per-episode human language instructions (from the DROID aggregated annotations). Only episodes with a non-empty instruction are included.

Caption ↔ video correspond by construction (same episode folder; verified by task_id / uuid match).

Selection criteria (episode kept only if complete)

RH20T: has caption + transformed/joint.npy + a color.mp4 for every camera view (the force/tactile sensor f0461559 is not counted as a camera). 2,862 episodes kept.
DROID: has trajectory.h5 + ≥3 camera MP4s + a non-empty language instruction. 39,986 kept (dropped mostly for missing caption — ~54% of DROID episodes carry a language instruction).

AgiBot usable lists

AgiBotWorld-Beta and AgiBotWorld2026 are not included here. The two *_usable_episodes.txt files list the episodes whose forward-kinematics skeleton projected correctly (complete render output); use them to filter the original datasets. Beta lines are task/episode; 2026 lines are Domain/Scene/task/range/ep.

Counts

Dataset	Episodes (complete / usable)
RH20T	2,862
DROID	39,986
AgiBotWorld-Beta (list only)	146,771
AgiBotWorld2026 (list only)	19,329