Blame for README.md

---
license: apache-2.0
language:
- en
- zh
task_categories:
- robotics
tags:
- robotics
- robot-learning
- bimanual-manipulation
- video
- webdataset
size_categories:
- 100K<n<1M
configs:
- config_name: agilex
  data_files:
  - split: train
    path: data/agilex/train/*.tar
  - split: validation
    path: data/agilex/validation/*.tar
- config_name: franka
  data_files:
  - split: train
    path: data/franka/train/*.tar
  - split: validation
    path: data/franka/validation/*.tar
- config_name: ur5
  data_files:
  - split: train
    path: data/ur5/train/*.tar
  - split: validation
    path: data/ur5/validation/*.tar
---

# RoboMIND 2.0 processed episodes

A processed subset of [RoboMIND 2.0](https://modelscope.cn/datasets/X-Humanoid/RoboMIND2.0)
containing **134,537 episodes** and approximately **846 hours** of bimanual robot demonstrations.
This is a derivative release, with credit to the original RoboMIND authors.

| Robot | Training episodes | Validation episodes |
| --- | ---: | ---: |
| AgileX | 19,027 | 2,084 |
| Franka | 60,069 | 6,616 |
| UR5 | 42,005 | 4,736 |
| Total | 121,101 | 13,436 |

The subset contains source-labeled successful demonstrations with structurally valid video and state/action streams.
Identical recorded payloads are deduplicated; alternate task instructions are retained in
`source_aliases` when present. Episodes with missing, incompatible, or frozen camera streams
are excluded. This release does not contain every embodiment or episode in the upstream dataset.

## UR5 action-quality caveat

The [Qwen-RobotManip technical report, Section 2.4](https://arxiv.org/html/2606.17846#S2.SS4)
reports that 81% of the RoboMIND UR-type episodes in its curation corpus failed a
state-action trend-alignment check. The report covers both RoboMIND and RoboMIND 2.0;
it does not identify which episodes in this derivative release are affected.
**The reported 81% is not a measured failure rate for this release.**

An exploratory audit on October 22, 2026 checked all **46,741 exported UR5 episodes**
and flagged **12 (0.026%)** with at least one moving arm joint below 0.70 directional
agreement after lag alignment. At thresholds 0.60 and 0.65, 7 and 8 episodes were
flagged respectively. Two of the 12 also had a joint whose best lag was negative.
The median estimated joint lag was 0.278 seconds, with actions preceding states.

This diagnostic used five-frame triangular smoothing, a lag search of approximately
plus or minus one second using position correlation, and agreement of the signs of
lag-aligned first differences. It scored arm joints only when both streams spanned
at least 0.05 radians, ignored differences at or below 0.00001 radians per frame,
and required at least ten scored differences. Grippers and low-motion joints were
excluded. A separate unsmoothed check of 512 reproducibly sampled UR5 episodes
flagged none at the 0.70 threshold. Twenty sampled episodes and all twelve flagged
episodes matched the original HDF5 values after the documented format conversion.

**This is not an exact reproduction of the paper's filter or a certification of
training quality.** It does not validate video-state alignment, kinematics, task
labels, or controller suitability. The recorded actions remain unchanged, without
temporal correction. We recommend excluding the following twelve episodes from
action-supervised training pending individual review, and assessing the remaining
UR5 data for the intended training setup.

<details>
<summary>UR5 episode keys flagged for review</summary>

```text
episode-17b9a9076462f3968e3455400ccd6d87f67b688294221a8bc274eb2a46ce594a
episode-2ddd2e1082689571a038f343359f8a7f5f8917b2ca6659edc8e36f50e5a0a237
episode-32eb11fbaeecd0ad13cf08974126d48d30e5efaaf17aa209cb38d9674df3db31
episode-4ba3a75369573c1cd79664131a70dd27566624dcfecbdb023c79ce93e19e7c09
episode-4d804fcf7fd753f39626a67daf75b96f4701cb906cf46c651713a91a95bab3ff
episode-5f6097f7ac3569a4282e9bf0b2fe3cc4ec5bf2008b02c4f8d127754791a23322
episode-7270a65a8afe8eab94b7bcd83866e531f4815b1d60e51cfe6beaceec60a89a71
episode-818172b46a58b29ba874da1d7765d084e3087ddfe66a7865dd4f07843ebc97d0
episode-8a48c96c88c8b345a10404db81f74efe3cde1d9bd0f72498750955c900ebcd2d
episode-a6d8d7020fd3f86643625c34ddcf9150bdf26164b114880c7b17c8470478f256
episode-d002afc3996d3e513062e129ed1d47793b8a8868706f5d0da8a20711fae483f8
episode-f2b1c676c204368342ea930dc7af795e2e332f911212d4873a36b592419f5269
```

</details>

## Format

Files are grouped into uncompressed WebDataset TAR shards of approximately 4 GiB under
`data/<robot>/<split>/`. Each episode has a stable key and three members:

- `<key>.mp4`: H.264 RGB video, with top, left-wrist, and right-wrist views stacked vertically.
  Each view is letterboxed to 640 × 480, giving a combined frame of 640 × 1440.
- `<key>.bin`: contiguous, row-major, little-endian float64 state/action values.
- `<key>.json`: task instructions, frame rate, frame count, column layout, camera layout,
  source episode identifiers, and other data interpretation metadata.

Video frame rates are capped at 30 fps, retaining lower native rates. One binary row corresponds
to one video frame. The exact frame rate and retained source frame indices are provided per episode.
AgileX and UR5 have 28 columns per row; Franka has 32. Read `row_layout.columns` for the exact order.
Robot joint values are in radians. Grippers are normalized to 0 = closed and 1 = open.
States describe the robot arms and actions describe the recorded leader arms, without added lag.
Camera calibration is unavailable and represented as `null`.

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"], metadata["row_layout"]["floats_per_row"]
)
```

The archive members contain the exported video and state/action bytes. Episode metadata is
reduced to public data fields. `dataset_info.json` summarizes the release and `shards.jsonl`
lists archive sizes, episode counts, and SHA-256 checksums.

## Quality-control audit

A full audit of **134,537 episodes** and **127,677,339 camera frames**
is published in [`audit/`](audit), covering the retained processed frames at revision
`72877243ff25e5f484ea216a1d684b65d1c0d7d0`. It includes joint, video, pose-continuity, workspace,
tracking/lag and duplicate checks, with explicit applicability limits.

**35,070 episodes triggered a configured threshold**, including
**232 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. RoboMIND native-pose/FK disagreements are review-only; published EEF poses use nominal joint FK. 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.

All RoboMIND EEF states and commands are calculated separately from the published
puppet and master joint rows. Franka uses `panda_link0` to `panda_link8`; UR5e uses
the manufacturer's nominal UR5e DH base to `tool0`; AgileX uses the dataset authors'
`aloha_new.urdf`, verified byte-identical to the manufacturer model, from each
`fl_base_link`/`fr_base_link` to `fl_link6`/`fr_link6`. The two arms' local base frames
are distinct; no missing common-base or camera calibration is fabricated.

The earlier UR5e rotation encoding discrepancy is avoided by reconstructing poses
from the corresponding joints after diagnosis. The twelve command-review episodes
listed above have command supervision masked out. These are nominal-model poses,
not factory-calibrated measurements or a certificate of camera reprojection accuracy.
Invalid AgileX native EEF fields are not used as ground truth or fitted calibration.

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="command"
)
# 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.

## Source and license

Upstream sources are [X-Humanoid/RoboMIND2.0-Agilex](https://modelscope.cn/datasets/X-Humanoid/RoboMIND2.0-Agilex),
[RoboMIND2.0-Franka](https://modelscope.cn/datasets/X-Humanoid/RoboMIND2.0-Franka-Part-1)
(parts 1–5), and [RoboMIND2.0-UR5](https://modelscope.cn/datasets/X-Humanoid/RoboMIND2.0-UR5).
They are distributed under Apache License 2.0, retained for this processed release; see `LICENSE`.
Changes include video encoding and resizing, temporal sampling, state/action formatting,
gripper normalization, filtering, deduplication, and dataset splits.

Please cite the original work:
Chengkai Hou et al. (2025), [RoboMIND 2.0: A Multimodal, Bimanual Mobile Manipulation Dataset
for Generalizable Embodied Intelligence](https://arxiv.org/abs/2512.24653).