The Dataset Viewer has been disabled on this dataset.

Zerith-H1-1K scene diversity

Zerith-H1-1K

English | 简体中文

Zerith-H1-1K is a thousand-hour real-world robot dataset for multi-task, long-horizon embodied manipulation, with fine-grained subtask segmentation and language annotations across 356 task configurations spanning sorting, office, kitchen, and dining scenarios.

Zerith-H1-1K brings together approximately 1,129 hours of real-world robot demonstrations across 356 task configurations, all collected on Zerith's own H1 robot platform. The dataset spans five major task categories: office document organization, tableware arrangement and storage, food storage and kitchen organization, target-object picking from bins, and bimanual object manipulation. Variations in task instructions, target objects, and manipulation requirements capture diverse robot manipulation processes in practical settings, providing real-world data at scale for imitation learning and multi-task policy learning.

Each demonstration includes RGB video from one high-mounted camera and two wrist-mounted cameras, along with robot actions, observed states, and Chinese task descriptions, recording both task semantics and the corresponding manipulation process. The dataset also provides language timelines and subtask interval fields. Some demonstrations include fine-grained subtask segments, supporting further research into the relationships among language instructions, manipulation stages, and action sequences.

The dataset uses the LeRobot v3.0 format and is organized and distributed by task, with each task archive independently loadable. Researchers can selectively download tasks according to their research goals, inspect observations, actions, and annotations, and build training pipelines for studies ranging from single-task imitation learning to multi-task policy learning.


1. Episode Examples

Four selected episodes illustrate multi-stage manipulation with synchronized left-wrist, head, and right-wrist views. Bilingual subtitles identify the active subtask; alternating blue and sand timeline segments show its duration.

Episode 1 · Bottle-cap tightening and storage

Episode 2 · Document organization

Episode 3 · Frozen-food storage

Episode 4 · Placemat and tableware preparation


2. Dataset Statistics

Property Value
Public task configurations 356
Frames 121,925,788
Recorded duration Approximately 1,128.94 hours
Compressed archive size 2.238 TB
Sampling rate 30 FPS
RGB camera streams 3
Dataset format LeRobot v3.0
Annotation language Chinese

3. Task Coverage

The release covers five broad task groups.

Task group Configurations Public task IDs Representative activities
Office Document Organization 23 task-000001–task-000023 Sort manuals and technical documents; archive project files; organize contracts, invoices, and receipts.
Tableware Arrangement and Storage 64 task-000024–task-000087 Set tables; collect used tableware; sort bowls and cutlery; return dishes to drawers and cabinets.
Food Storage and Kitchen Organization 90 task-000088–task-000177 Store and sort food in refrigerators and freezers; organize condiments; arrange storage containers.
Target-Object Picking from Bins 150 task-000178–task-000327 Pick a target product from a bin, with variations including packaged snacks, candy, milk, and beverages.
Bimanual Object Manipulation 29 task-000329–task-000357 Pick up two different tabletop objects simultaneously with both hands; one configuration also includes placing them into bins on either side.

4. Download and Quick Start

4.1 Download One Task

Install the Hugging Face CLI. This dataset is public and can be downloaded without logging in.

pip install -U huggingface_hub

hf download OpenZerith/Zerith-H1-1K \
  --repo-type dataset \
  --include 'data/task-000286/*' \
  --local-dir ./Zerith-H1-1K

Extract this single-archive task into a separate directory:

unzip ./Zerith-H1-1K/data/task-000286/task-000286.zip \
  -d ./extracted/task-000286-single

4.2 Download the Full Release

Downloading the full release requires approximately 2.238 TB of storage for the compressed archives, plus additional space for extraction.

hf download OpenZerith/Zerith-H1-1K \
  --repo-type dataset \
  --local-dir ./Zerith-H1-1K

4.3 Load an Extracted Dataset

Install LeRobot 0.7.0 and TorchCodec. Python 3.12 or later and FFmpeg shared libraries are required.

pip install 'torch==2.10.0' 'torchvision==0.25.0' 'torchcodec==0.10.0' --index-url https://download.pytorch.org/whl/cpu
pip install 'lerobot[dataset]==0.7.0'
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="OpenZerith/task-000286",
    root=Path("./extracted/task-000286-single/task-000286"),
    video_backend="torchcodec",
)

sample = dataset[0]
print(sample["task"])
print(sample["action.end_0_pose"].shape)
print(sample["observation.images.cam_high_0"].shape)

Set root to the extracted directory containing meta/, data/, and videos/. Here, repo_id is a local identifier.


5. Repository and Archive Structure

Zerith-H1-1K/
├── README.md
├── README_zh-CN.md
├── assets/
│   ├── zerith_h1_scene_diversity.png
│   └── episodes/          # Example videos and preview images
└── data/
    ├── task-000001/
    │   └── task-000001.zip
    ├── task-000009/
    │   ├── task-000009-part-000.zip
    │   └── task-000009-part-001.zip
    ├── ...
    └── task-000357/
        └── task-000357.zip

Three tasks have two archives each: task-000009, task-000038, and task-000045. Each archive is an independently loadable dataset, with its own metadata and local episode and frame indices. Download all archives to obtain all data for one of these tasks.

Inside one archive:

task-XXXXXX/
├── meta/
│   ├── info.json
│   ├── tasks.parquet
│   └── episodes/chunk-000/file-000.parquet
├── data/
│   └── chunk-000/file-000.parquet
└── videos/
    ├── observation.images.cam_high_0/chunk-000/file-000.mp4
    ├── observation.images.cam_wrist_0/chunk-000/file-000.mp4
    └── observation.images.cam_wrist_1/chunk-000/file-000.mp4

Chunk and file counts vary by archive. Unlike an episode-per-file layout, LeRobot v3.0 can store multiple episodes in a Parquet or MP4 shard. Episode metadata supplies the boundaries and references needed to reconstruct each episode.


6. Data Fields

6.1 Metadata and Indexing

File Contents
meta/info.json Format version, robot type, FPS, episode/frame counts, splits, feature definitions, and data/video path templates.
meta/tasks.parquet Original task descriptions and their local task_index values.
meta/episodes/chunk-*/file-*.parquet Episode lengths, task descriptions, global frame ranges, and references into data and camera shards.
data/chunk-*/file-*.parquet One row per frame: numeric actions/states, time and indices, language timeline, and subtask interval fields.
videos/<camera>/chunk-*/file-*.mp4 RGB camera recordings shared across episodes within a shard.

The frame table has 27 physical columns. Three additional video features are described in info.json and stored as MP4 files, giving 30 feature definitions. Always inspect the selected archive's info.json for its actual shapes and video metadata.

6.2 Actions and Observed States

The numeric fields below use float32.

Action field Observed-state field Shape of each field Components
action.end_0_pose, action.end_1_pose observation.state.end_0_pose, observation.state.end_1_pose [7] Position [x,y,z], followed by quaternion [x,y,z,w].
action.arm_0_joint_state, action.arm_1_joint_state observation.state.arm_0_joint_state, observation.state.arm_1_joint_state [7] Seven joint position components per arm.
action.effector_0_joint_state, action.effector_1_joint_state observation.state.effector_0_joint_state, observation.state.effector_1_joint_state [1] One end-effector position component.
action.waist_joint_state observation.state.waist_joint_state [3] Position [up-down,yaw,pitch].
action.head_joint_state observation.state.head_joint_state [2] Position [yaw,pitch].
action.base_joint_state observation.state.base_velocity [2] The action field name describes wheel positions; the observed-state field name describes wheel velocities. These are different quantities.

6.3 Cameras

Feature View
observation.images.cam_high_0 High-mounted camera
observation.images.cam_wrist_0 Wrist camera 0
observation.images.cam_wrist_1 Wrist camera 1

The release uses H.264 RGB video at 30 FPS, typically 640 × 480. Tasks task-000080 and task-000352 retain 848 × 480 recordings.

6.4 Frame Indices and Language

Field Type Meaning
timestamp float32 Episode-relative time in seconds on the fixed-FPS timeline.
frame_index int64 Zero-based frame number within the episode.
episode_index int64 Episode number within this independent archive.
index int64 Frame index across the archive's dataset.
task_index int64 Reference to the local task-description dictionary.
language_persistent Native language column Episode language timeline, represented as structured language atoms.
subtask_segment_index int64 Current segment number; -1 for an unannotated frame.
subtask_start_frame int64 Inclusive segment start; -1 when unannotated.
subtask_end_frame int64 Exclusive segment end; -1 when unannotated.

6.5 Language and Subtask Annotations

Language atoms have six fields: role, content, style, timestamp, camera, and tool_calls. content contains the subtask description. timestamp specifies when the description becomes active, in seconds relative to the start of the episode.

Segment boundaries use episode-local frame intervals [start, end). For example, [10, 20) includes frames 10 through 19.


7. Usage Notes and Limitations

  • Normalization: Inspect the archives you use and compute any missing normalization statistics from your training split.
  • Training adaptation: Actions and states are stored in separate named fields, rather than a single concatenated vector. A training pipeline must explicitly select and arrange the features it consumes. The loading example does not guarantee compatibility with every policy's default training configuration.
  • Splits: Define and document the training and evaluation splits used in your experiments.
  • Versioning: Record the Hub revision, archive names, selected episodes, loader version, and preprocessing configuration for reproducibility.

8. License and Citation

This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

You may share and adapt the licensed material subject to the license terms, including attribution, a link to the license, and an indication of modifications. Commercial use is not permitted under this license. Adapted material that you distribute must follow the ShareAlike terms. See the official license summary and legal text for the governing terms.

If you use this dataset, please cite the following technical report:

@techreport{zerith2026zerith25,
  title       = {ZERITH 2.5: From Predictive Power to Policy Learnability},
  author      = {{ZERITH Z-Lab}},
  institution = {ZERITH Z-Lab},
  year        = {2026},
  month       = nov
}

Please include the repository revision used in your experiment records. Questions and dataset issues can be raised through the repository's Community page.

Downloads last month
3,593