README.md

---
license: cc-by-nc-sa-4.0
language:
- en
task_categories:
- robotics
tags:
- robotics
- robot-learning
- bimanual-manipulation
- video
- webdataset
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/agibot/train/*.tar
  - split: validation
    path: data/agibot/validation/*.tar
---

# AgiBot World Beta processed episodes

A processed subset of [agibot-world/AgiBotWorld-Beta](https://huggingface.co/datasets/agibot-world/AgiBotWorld-Beta),
containing **142,685 verified episodes**, approximately **2,330.1 hours**, and **6.22 TB** of exported data.
Credit for the original demonstrations and annotations belongs to the AgiBot World contributors.
This is a derivative release.

| Split | Episodes | Hours |
| --- | ---: | ---: |
| Training | 128,530 | 2,098.7 |
| Validation | 14,155 | 231.4 |

This subset requires all three camera streams, robot state/action data, camera intrinsics,
and task labels. Episodes with more than 1 cm of base translation are excluded, as are
three source episodes with unreadable head-camera video. All retained source frames are preserved.
The source dataset contains additional episodes and modalities outside this processed subset.
The source annotations do not provide episode success/failure labels.

## Format

Uncompressed WebDataset TAR shards of approximately 4 GiB are stored under
`data/agibot/train/` and `data/agibot/validation/`. Each episode has a stable key and three members:

- `<key>.mp4`: H.264 video at 30 fps. The head, left-wrist, and right-wrist camera views
  are each resized to 640 × 480 and stacked vertically into a 640 × 1440 frame.
- `<key>.bin`: contiguous, row-major, little-endian float64 state/action values.
- `<key>.json`: task instructions, scene description, subtask annotations, frame counts,
  column layout, camera intrinsics, and public source episode identifiers.

Each binary row corresponds to one video frame. The 32 columns contain left-arm state
(7 joints and gripper), right-arm state (7 joints and gripper), and the corresponding
left/right action values. The exact order is provided in `row_layout.columns`.
Grippers are normalized to 0 = closed and 1 = open. Joint values are in radians.
Source joint-action values match the joint-state values and are preserved as recorded;
gripper actions use the recorded commands. No additional action lag is introduced.
Camera intrinsics describe the individual resized views; see per-episode metadata.

For an extracted episode:

```python
import json
import numpy as np

with open("<key>.json", encoding="utf-8") as stream:
    metadata = json.load(stream)
rows = np.fromfile("<key>.bin", dtype="<f8").reshape(metadata["num_steps"], 32)
```

`dataset_info.json` summarizes the release. `shards.jsonl` provides archive sizes,
episode counts, and SHA-256 checksums. Episode metadata contains only public data fields.

## Quality-control audit

A full audit of **142,685 episodes** and **754,950,021 camera frames**
is published in [`audit/`](audit), covering the retained processed frames at revision
`6fc8f88d74250c492622e7ff03586e162b852b2f`. It includes joint, video, pose-continuity, workspace,
tracking/lag and duplicate checks, with explicit applicability limits.

**96,315 episodes triggered a configured threshold**, including
**1,450 EEF-state spike candidates**. Broad acceleration/jerk
flags are heuristic, and dark-frame detections can include camera occlusion. These
are **review candidates, not an automatic removal list or demonstration outcome labels**.
Check counts overlap; [`audit/summary.json`](audit/summary.json) records the full breakdown.

- [`audit/episode_flags.csv`](audit/episode_flags.csv): public episode IDs, robot, split and check reasons.
- [`audit/ids/`](audit/ids): aggregate and per-check ID lists, including separate EEF state/command lists.
- [`audit/measurements/`](audit/measurements): gzip JSONL measurements for every episode and audit pass.
- [`audit/methods/`](audit/methods): thresholds and training-only statistical workspace bounds.
- [`audit/duplicate_groups.jsonl`](audit/duplicate_groups.jsonl): canonical/duplicate IDs and cross-split indicators.
- [`audit/files.jsonl`](audit/files.jsonl): public audit file sizes and SHA-256 checksums.

Measurement JSON uses `flagged_checks`; CSV check names are separated by semicolons.
The CSV contains episodes with a flag or review annotation; measurement files cover
every episode. Frame indices are zero-based retained video indices. EEF spike
measurements also give source indices and affected `next_state` action indices.
The native-pose/FK comparisons refer to upstream recorded fields, separate from
the published nominal-FK targets. Copied joint commands cannot establish independent
tracking quality. Independent AgiBot FK could not be checked without a verified robot model. Physical calibration and dataset-wide camera reprojection remain unverified.

**No episodes, videos, joint binaries or EEF payloads were removed or repaired.**
The published EEF arrays retain the audited values, including flagged anomalies.

## EEF action/state companions

The optional `eef/` companion release covers every episode in this processed subset,
using the same stable `episode_id` and exactly the same retained video frames. The
original joint binaries and videos remain available. Companion TARs contain
`<episode_id>.npz` arrays and `<episode_id>.json` frame/pose provenance.

`state_pose` is `[T, 2, 7]` float32: XYZ in metres followed by quaternion **xyzw**,
left arm then right arm. `command_pose` uses the same layout; when a recorded command
stream exactly copies the state, the array is stored once and the metadata declares
the alias. Gripper arrays use **0 = closed, 1 = open**. Boolean pose and gripper masks
must be respected. `source_frame_index` preserves the video-to-source row mapping.

Select action semantics explicitly in `eef/reader.py`:

- `action_mode="command"`: the corresponding recorded-command pose, with no added
  timestep shift. Copied source pose commands are excluded from supervision by default.
- `action_mode="next_state"`: the achieved state at the next retained frame, including
  gripper state. This is a derived target, not an original controller command. The
  final target is masked; do not shift it again in another loader.

AgiBot retains the source robot/ego-frame flange state poses. All 142,685 episodes
have valid finite state-pose arrays. Native EEF commands exactly copy state poses
in 141,318 episodes. Of the remaining episodes, 1,364 have shorter command streams
and three have distinct dense commands. Sparse commands are aligned only at valid
source indices; the fifty episodes with unresolved indices have pose-command masks
set to false. Uncommanded rows are not silently padded with state poses.

For the OpenWAM-Alpha future-state recipe, select `action_mode="next_state"`.
`command` mode preserves the native command distinction and its missing-data masks;
it cannot recover independent commands from the source copies. Gripper commands
remain separate. The exact URDF root origin and dataset-wide visual projection
accuracy have not been independently certified.

The reader emits `[T, 80]` state/action arrays and masks by default, with the left
20-D schema arm in dimensions `0:10` and the right arm in `34:44`; the unused slots
are masked. Set `unify_action=False` for `[T, 20]`: each arm is XYZ + the first two
rotation-matrix columns (6D rotation) + gripper. Poses use the frames named in each
episode, not camera coordinates or a universal fingertip TCP.

From a downloaded dataset root, join an existing episode's metadata to its companion:

```python
from eef.reader import load_index, load_indexed_episode

eef_index = load_index("eef/episodes.jsonl")
targets = load_indexed_episode(
    ".", eef_index[metadata["episode_id"]], action_mode="next_state"
)
# targets["state"], targets["action"], targets["state_mask"], targets["action_mask"]
```

`eef/episodes.jsonl` supplies direct TAR offsets. `eef/shards.jsonl` supplies file
sizes and SHA-256 checksums; `eef/info.json` records full coverage and verification.
`eef/models.json` documents frames, kinematics and limitations. `eef/statistics.json`
contains separate state, command and next-state statistics from the **training split
only**; stored poses are unnormalized. Keep rotation-6D coordinates unnormalized.

The portable reader matches the OpenWAM pose layout; integrating these companions
with a training pipeline's video/window reader is still required. The original
OpenWAM LeRobot reader does not directly consume these WebDataset companions.

## License and attribution

This processed release retains the upstream **Creative Commons Attribution-NonCommercial-ShareAlike
4.0 International (CC BY-NC-SA 4.0)** license. See `LICENSE` and the
[license terms](https://creativecommons.org/licenses/by-nc-sa/4.0/).
Noncommercial use, attribution, and sharing adaptations under the same license are required.

The original source is [AgiBotWorld-Beta](https://huggingface.co/datasets/agibot-world/AgiBotWorld-Beta)
from the [AgiBot World contributors](https://github.com/OpenDriveLab/AgiBot-World).
Changes in this release include video resizing and encoding, state/action formatting,
gripper normalization, selection of stationary-base episodes, dataset splits, metadata reduction,
and archive packaging.

Please cite the original project: AgiBot World Colosseum contributors (2024),
*AgiBot World Colosseum*, https://github.com/OpenDriveLab/AgiBot-World.