airoa-org/airoa-pi05-hsr-base
Robotics • Updated • 2
一般公開向けデータセット利用規約への同意
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
一般公開向けデータセット利用規約
一般公開向けデータセット利用規約は、一般社団法人AIロボット協会が提供する学習用データセットの利用に関して、利用者が遵守すべき利用条件を定めるものです。利用者は、一般公開向けデータセット利用規約に同意し、又は同データセットを利用することにより、一般公開向けデータセット利用規約に定めるすべての条件に法的に拘束されることに同意するものとします。
第1条 定義
一般公開向けデータセット利用規約内の各用語は、以下の意味を有するものとします。
(1) 「本契約」とは、利用者と当協会の間で成立する、一般公開向けデータセット利用規約を内容とする契約を意味します。
(2) 「当協会」とは、一般社団法人AIロボット協会を意味します。
(3) 「一般公開共通データセット」とは、Phase1で収集されたHSRデータのクレンジング済みデータのうち、当協会が一般公開共通データセットとして指定し、当協会で指定したHugging FaceのURLにアップロードする方法で開示された学習用データセット(5000時間分)を意味します。
(4) 「利用者」とは、一般公開共通データセットを利用する個人又は法人を意味します。
(5) 「ライセンス」とは、当協会が利用者に対して、本契約の条件に基づき一般公開共通データセットを利用することを許諾することを意味します。
(6) 「派生データセット」とは、一般公開共通データセットを含む学習用データセット及び一般公開共通データセットを基に作成された学習用データセットを意味します。
(7) 「派生モデル」とは、一般公開共通データセットを用いて作成されたAIモデルを意味します。
(8) 「派生モデル等」とは、派生モデル及び派生データセットを意味します。
第2条 ライセンス
当協会は、利用者に対し、本契約の条項に従い、かつ本契約に規定された範囲内で、一般公開共通データセットを利用する許可を付与します。ただし、利用者がこれらのすべての条項に同意し、遵守することを条件とします。
本契約により提供されるライセンスは、非独占的、全世界的、取消可能、譲渡不可、及びロイヤリティ無料とします。
利用者は、一般公開共通データセットを、本契約に定める条件で利用することができるものとします。
第3条 データセットの利用条件
利用者による一般公開共通データセットの利用は、次の各号に定めるところによるものとします。
(1) 利用者は、一般公開共通データセットを、自らの研究開発の目的に限り使用することができ、第三者に対する有償販売その他の方法により利用してはならないものとします。
(2) 利用者は、自らの研究開発の目的で、一般公開共通データセットを用いて、派生モデルの学習を行い、又はこれらのデータセットを加工することができるものとします。
(3) 利用者は、派生モデルを自らの研究開発の目的に限り使用することができるものとし、第三者に再配布してはならないものとします。
(4) 利用者は、一般公開共通データセット、派生データセット及びその複製物を第三者に再配布してはならないものとします。
第4条 派生モデル等及び出力の提供
利用者は、自らが作成した派生モデル等及び派生モデルからの出力を当協会に提供するよう努めるものとします。
前項に基づいて利用者が当協会に提供した派生モデル等及び派生モデルからの出力について、当協会は、商用・非商用問わず、かつ用途の制限なく利用することができるものとします。なお、利用には、派生モデル等及び同派生モデルからの出力の改変、複製、再配布並びに派生モデル等及び派生モデルからの出力を用いたサービスの実施を含みます。
第5条 保証の否認
一般公開共通データセットは「現状有姿」で提供されます。当協会は、それらに関して、正確性、真正性、商品性、品質、性能、特定用途への適合性、又は権利の非侵害性を含み（これらに限定されません。）、一切の保証をしません。利用者は、自らの責任において一般公開共通データセットの利用の適切性を判断するものとし、一般公開共通データセットの利用の結果生じるあらゆる結果について、利用者は全責任を負うものとします。
第6条 責任の制限
一般公開共通データセットに関して利用者及び第三者が被った損害に対して、契約、不法行為、製造物責任、その他の法的請求の根拠の如何を問わず、当協会は一切の責任を負わないものとします。
第7条 利用者の責任
利用者は、一般公開共通データセットの取得及び利用に際して、本契約の条件並びにすべての関連法規（個人情報の保護に関する法律、著作権法、外国為替及び外国貿易法を含むがこれに限られない。）に準拠しなければなりません。
利用者は、本契約の違反、又は一般公開共通データセットの利用により当協会に発生した損害を補償しなければなりません。
利用者が一般公開共通データセットを利用したことにより当協会が第三者から損害賠償請求その他の請求を受けた場合には、利用者は自己の責任と費用負担によって当該請求を免れさせ、当協会に何らの損害も被らせないものとします。
第8条 権利の帰属
一般公開共通データセットに関する知的財産権を含む一切の権利（以下、本条で「知的財産権等」といいます。）は、当協会に帰属します。
利用者が派生データセットを作成する際に一般公開共通データセットに対して行った変更部分に関する知的財産権等は、当該変更を行った利用者に帰属します。
第9条 本契約の有効期間
本契約は、利用者が本契約に同意した時点、又は一般公開共通データセットにアクセスした時点のいずれか早い時点で発効し、本契約の条項にしたがって終了するまで有効に存続するものとします。
利用者が本契約の各条項に違反した場合、当協会は何らの催告なく本契約を終了することができます。
第10条 本契約の改定
当協会は、本契約をみずからの裁量で改訂することができます。当協会は、変更内容及び変更の効力発生時期を含め、本契約の改訂内容を、変更の実施前に、当協会所定の方法により公表します。
第11条 準拠法及び管轄裁判所
(1) 本契約は、日本法に準拠します。
(2) 本契約又は一般公開共通データセット、又は派生データセットに起因又は関連して生じる紛争は、東京地方裁判所をもって第一審の専属合意管轄裁判所とします。
第12条 言語
本契約は日本語を正本とします。
2026年7月31日制定
Log in or Sign Up to review the conditions and access this dataset content.
AIRoA MoMa 5k
AIRoA MoMa 5k is a large-scale, task-structured dataset of real-robot mobile manipulation collected by teleoperating Toyota Human Support Robots (HSRs). The public release contains 1,184,259 successful Primitive-Action (PA) episodes, 180,905,084 frames, and 5,025 recorded hours from 44 physical robots at five collection sites.
Each PA remains independently addressable for policy training, while execution-level metadata preserve the Short-Horizon Task (SHT) in which that PA occurred and its position in the ordered PA sequence. The release combines head- and hand-camera RGB, robot state, mobile-base actions, wrist force-torque history, end-effector pose, low-level servo telemetry, and task hierarchy metadata in LeRobot v3.0 format.
The repository is published in the packed LeRobot v3.0 layout, which is what
LeRobotDataset loads. The original per-episode v2.1 files remain available in
the companion Storage Bucket (see Download).
Scope of this card. This card describes the public, success-only AIRoA MoMa 5k artifact with release identifier
airoa-moma-5k.
Dataset summary
| Property | AIRoA MoMa 5k |
|---|---|
| PA episodes | 1,184,259 |
| Frames | 180,905,084 |
| Recorded duration | 5,025.1 h at 10 Hz |
| Reconstructed SHT instances | 425,518 |
| SHT templates | 93 |
| Defined / observed PA labels | 531 / 358 |
| Distinct PA sequences | 1,985 |
| Sites / physical HSR units | 5 / 44 |
| Camera views | 2 (head and hand) |
| Outcome | Successful episodes only |
| Split | Train only |
| Storage | Approximately 4.0 TiB |
| Format | LeRobot v3.0; Parquet + AV1 video |
| Files | 13,318 (4,840 data shards, 8,080 video shards, 398 metadata) |
| Collection period | April 2025 to January 2026 |
| Release identifier | airoa-moma-5k |
The 5k release expands the preliminary AIRoA MoMa dataset (25,469 PA episodes, approximately 94 h, one site) by 46.5 times in episodes, 53.5 times in recorded hours, and five times in collection sites.
Task organization
Primitive Actions and Short-Horizon Tasks
The unit stored as an episode is a Primitive Action (PA), such as navigating to a container, grasping an object, or placing it on a shelf. Multiple PA episodes can be grouped into a Short-Horizon Task (SHT) execution.
tasks[0]is the natural-language instruction for the active PA episode.short_horizon_taskis the parent SHT template.primitive_actionstores the parent task's ordered PA sequence.uuidis the SHT execution identifier shared by PA episodes from the same execution.primitive_action_indexin the frame-level Parquet data identifies the PA's position in that sequence.task_indexindexes the PA instruction vocabulary; it is not the sequence position.
Grouping by uuid reconstructs 425,518 SHT instances with a mean of 2.78 released PAs per instance. The public core covers 93 SHT templates: 68 main task templates, which account for 99.6% of released episodes, plus 25 site-specific template variants.
The distribution is long-tailed. The most frequent SHT contains 71,258 PA episodes, and 173 labels in the defined 531-label PA vocabulary are not observed in this release.
Duration
| Unit | Median | 95th percentile | Maximum |
|---|---|---|---|
| PA episode | 12.4 s | 36.3 s | 280 s |
| Reconstructed SHT instance | 31.7 s | 115.9 s | 922 s |
Collection distribution
All data were collected with the Toyota Human Support Robot using human leader-follower teleoperation. The public core contains data from five sites:
| Site | PA episodes | Share |
|---|---|---|
| telexistence | 719,313 | 60.7% |
| kyutech | 362,934 | 30.6% |
| weblab | 64,050 | 5.4% |
| fastlabel | 33,354 | 2.8% |
| heiwajima | 4,608 | 0.4% |
Site and robot provenance are retained per episode. Pseudonymous operator identifiers are available for the telexistence subset only, where 122 distinct operator identifiers are observed; operator identifiers are absent from the other four sites.
These counts describe data coverage, not task difficulty or policy performance. Users should account for the substantial site and task imbalance when constructing evaluation splits.
Data format
The release follows the LeRobot v3.0 layout: episodes are packed into size-bounded shard files rather than stored one file per episode.
meta/
├── info.json
├── stats.json
├── tasks.parquet
├── anomalous_episodes.json
└── episodes/
└── chunk-{chunk:03d}/
└── file-{file:03d}.parquet # per-episode metadata + stats
data/
└── chunk-{chunk:03d}/
└── file-{file:03d}.parquet # packed frame data
videos/
├── observation.image.head/
│ └── chunk-{chunk:03d}/
│ └── file-{file:03d}.mp4 # concatenated AV1, ~500 MB each
└── observation.image.hand/
└── chunk-{chunk:03d}/
└── file-{file:03d}.mp4
13,318 files in total: 4,840 data shards, 8,080 video shards and 398 metadata files.
Each episode's location inside a shard is given by its metadata row
(data/chunk_index, data/file_index, and per-camera
videos/<key>/{chunk_index, file_index, from_timestamp, to_timestamp}), which
LeRobotDataset resolves for you. Videos are stream-copy concatenations: the AV1
payload is bit-identical to the per-episode files in the v2.1 bucket.
The only provided split is:
train: 0:1184259
Core frame-level features
All streams are synchronized at 10 Hz.
| Feature | Shape / type | Description |
|---|---|---|
observation.image.head |
480 x 640 x 3 video | Head-mounted RGB, AV1, 10 fps |
observation.image.hand |
480 x 640 x 3 video | Hand-mounted RGB, AV1, 10 fps |
observation.state |
float32[8] | Arm, wrist, gripper, and head joint state |
observation.wrist.wrench |
float32[600] | Flattened history of 100 six-axis force-torque samples in the wrist-sensor frame; force in N and torque in N m. Exception: in 4,608 episodes (1,179,651-1,184,258) only a single six-value sample per frame was recorded, not the 100-sample history. In the v2.1 bucket these arrays are float32[6], not float32[600] - consumers reading v2.1 directly must handle both widths. In v3.0 they are padded to [600]; see Limitations. |
observation.end_effector_pose.absolute |
float32[6] | Absolute end-effector x, y, z, roll, pitch, yaw |
observation.end_effector_pose.relative |
float32[6] | Relative end-effector pose |
action.absolute |
float32[8] | Absolute arm, gripper, and head action representation |
action.state_diff |
float32[8] | Difference between next and current robot state |
action.relative |
float32[11] | Relative arm, gripper, head, and mobile-base action |
action.arm |
float32[5] | Arm action |
action.gripper |
float32[1] | Gripper action |
action.head |
float32[2] | Head action |
action.base |
float32[3] | Mobile-base x, y, theta delta |
command.servo.position |
float32[11] | Low-level servo position command |
command.servo.velocity |
float32[11] | Low-level servo velocity command |
state.servo.* |
float32[11] each | Motor position, driven position, velocity, temperature, effort, and current |
*.is_fresh |
bool arrays | Freshness indicators for streams whose values may be carried forward during synchronization |
The 8-D joint ordering is:
arm_lift_joint, arm_flex_joint, arm_roll_joint, wrist_flex_joint,
wrist_roll_joint, hand_motor_joint, head_pan_joint, head_tilt_joint
The 11-D relative-action ordering appends:
base_x, base_y, base_t
The servo features cover the eight joints above plus:
base_l_drive_wheel_joint, base_r_drive_wheel_joint, base_roll_joint
Additional standard indexing fields include episode_index, frame_index, timestamp, next.done, short_horizon_task_index, primitive_action_index, success_primitive_action, and task_index.
Episode metadata
In v3.0, per-episode metadata lives in meta/episodes/chunk-*/file-*.parquet (one row per episode, plus flattened per-episode statistics and the shard/timestamp pointers described above). In the v2.1 bucket the same records are one JSON object per line in meta/episodes.jsonl. The fields are otherwise the same; a shortened example is:
{
"episode_index": 0,
"tasks": ["Grab one 500-600 ml pet bottle"],
"length": 263,
"location_name": "telexistence",
"interface": "hsr_leader_teleop",
"hsr_id": "hsrb101",
"task_type": "PA",
"task_success": true,
"short_horizon_task": "Take one 500-600 ml pet bottle from the container and stock it on the shelf.",
"primitive_action": [
"Navigate to container",
"Grab one 500-600 ml pet bottle",
"Go back to shelf",
"Place 500-600 ml pet bottle on shelf",
"Return to charging station"
],
"success_short_horizon_task": true,
"uuid": "1d449b4e-4e7e-4321-bb69-a59bccfb31ee"
}
The complete records also contain collection-component provenance and source-version metadata. Do not assume that every optional provenance field is populated at every site.
Download
Two forms of the same data are published: the v3.0 repository (packed, loads with LeRobot) and the v2.1 Storage Bucket (original per-episode files).
v3.0 repository (recommended)
from lerobot.datasets.lerobot_dataset import LeRobotDataset
# a subset - downloads only the shards backing these episodes
ds = LeRobotDataset("airoa-org/airoa-moma-5k", episodes=[0, 1, 2])
# the whole dataset - downloads ~4.0 TiB
ds = LeRobotDataset("airoa-org/airoa-moma-5k")
A default load downloads the entire ~4.0 TiB. Pass
episodes=[...]to fetch only the shards you need. Access is gated: accept the terms on the dataset page and runhf auth loginfirst.
Resolution goes through the v3.0 tag, so no revision argument is needed.
v2.1 raw tree (Storage Bucket)
The v2.1 raw tree is approximately 4.0 TiB across more than 3.5 million files,
distributed through the Hugging Face Storage Bucket
airoa-org/airoa-moma-5k,
which supports efficient bulk transfer without per-file Hub API overhead. Use it for
per-episode access without LeRobot, or for bulk mirroring.
hf sync
Install the Hugging Face CLI and sync the bucket (or any prefix of it) to a local directory:
pip install -U huggingface_hub
# full mirror (~4.0 TiB)
hf sync hf://buckets/airoa-org/airoa-moma-5k ./airoa-moma-5k
# metadata only
hf sync hf://buckets/airoa-org/airoa-moma-5k/meta ./airoa-moma-5k/meta
# a single chunk (1,000 episodes: videos of both cameras)
hf sync hf://buckets/airoa-org/airoa-moma-5k/videos/chunk-000 ./airoa-moma-5k/videos/chunk-000
hf sync hf://buckets/airoa-org/airoa-moma-5k/data/chunk-000 ./airoa-moma-5k/data/chunk-000
hf sync compares source and destination and transfers only missing or changed files, so an interrupted download can be resumed by re-running the same command. For partial use, prefer prefix-scoped syncs as shown above rather than listing the full 3.5-million-object tree.
Anonymous downloads are subject to stricter per-IP rate limits; logging in with a free Hugging Face account token (hf auth login) is recommended for large transfers.
Alternative: S3-compatible API
The bucket is also reachable through an S3-compatible gateway, which suits streaming transfers into other storage systems with tools such as rclone, s5cmd, or the AWS SDKs:
- endpoint:
https://s3.hf.co/airoa-org(path-style addressing, ListObjectsV2, regionus-east-1) - bucket name:
airoa-moma-5k - credentials: generate S3 credentials from a Hugging Face access token (see the S3 compatibility documentation)
Object downloads redirect to the nearest CDN edge, so S3 clients that follow redirects (rclone, s5cmd, curl) download at CDN speed.
Programmatic access
huggingface_hub exposes bucket listing and per-file access (HfApi.list_bucket_tree, hf:// paths), which supports partial and streaming reads — for example querying the Parquet data directly with DuckDB without downloading videos.
The bucket is flat object storage and does not integrate with hub-style loading; for
LeRobotDataset, use the v3.0 repository above.
Collection, filtering, and privacy
SHT labels, PA labels, PA boundaries, ordered task structure, and outcome flags were recorded during data collection. Processing verifies and packages those annotations through:
- signal-validity and motion checks;
- segmentation at the recorded PA boundaries;
- removal of invalid, negligible-motion, duration-outlier, and signal-outlier segments;
- visual inspection of random samples; and
- bystander screening of both camera streams.
For the public release, a conventional person detector was run over both camera streams, with an explicitly controlled confidence threshold and false-positive trade-off. Videos with bystander detections above the release threshold were privacy-blurred rather than removed: 203,211 of the 2,368,518 videos (8.6%, spanning 175,767 episodes) are blur-processed renders. No episodes were dropped on this basis, and every episode's robot state, actions and force-torque data are unmodified.
Intended uses
AIRoA MoMa 5k is intended for research in:
- imitation learning and offline robot learning;
- vision-language-action model training and adaptation;
- mobile-manipulation perception and control;
- hierarchical or compositional task modeling;
- multi-view robot vision;
- contact-aware learning using wrist force-torque history;
- analysis of task, site, robot, and operator-domain variation.
The metadata support held-out-site splits and compositional splits that hold out complete SHT templates while retaining some constituent PA labels in training. When splitting by SHT, group by uuid so that PA episodes from one execution do not leak across splits.
Limitations and out-of-scope uses
Success-only release: Failed and suboptimal episodes from the broader AIRoA collection are not included in this public core. The dataset is unsuitable for failure-frequency estimation.
No standardized benchmark: The release provides a training corpus, not fixed policy inputs, train/test splits, success criteria, or a leaderboard.
Single embodiment: All episodes use HSR. Results may not transfer directly to other robot morphologies or controllers.
Long-tailed coverage: Tasks, PA labels, sites, robots, and operators are not balanced.
Templated language: Instructions are task templates rather than free-form operator utterances.
Incomplete operator provenance: Operator identifiers are present only for the telexistence subset.
Successful demonstrations are not safety guarantees: Policies trained on this dataset require independent validation, workspace constraints, collision handling, and human oversight before real-world deployment.
Padded force-torque history in 4,608 episodes: episodes 1,179,651-1,184,258 (all from the heiwajima site) were recorded with a single six-value force-torque sample instead of the 100-sample history. In the v3.0 release that sample is tiled 100x so the array matches the declared
float32[600]shape; the values are correct but the history dimension is synthetic - all 100 slots hold the current sample. Treat wrist-wrench temporal structure in these episodes as absent, not as measured. The episode list is published atmeta/anomalous_episodes.json.In the v2.1 Storage Bucket these episodes are not padded: each frame holds a
float32[6]array (the single sample) where every other episode holdsfloat32[600]. Code that assumes a fixed 600-wide wrench array across the v2.1 tree will fail on these 4,608 episodes - check the array width per episode, or use the v3.0 release where it is uniform.Known metadata edge case: In 95 of 1,184,259 episode records, the active
tasks[0]string is not found verbatim in the storedprimitive_actionsequence. Consumers should not require exact string membership without an explicit fallback.
The dataset should not be used to infer personal attributes, identify operators, or develop surveillance systems. Site, robot, and pseudonymous operator fields are provided for provenance and domain-shift research.
License
Use of this dataset is governed by the AIRoA Public Dataset Terms of Use (一般公開向けデータセット利用規約). The Japanese text in LICENSE.ja.md is the authoritative version (Article 12). Key conditions include: research and development use only, no commercial resale, and no redistribution of the dataset, derivative datasets, or derivative models to third parties.
Citation
The preliminary release is described in:
@article{takanami2025airoamoma,
author = {Takanami, Ryosuke and Khrapchenkov, Petr and Morikuni, Shu and others},
title = {{AIRoA MoMa} Dataset: A Large-Scale Hierarchical Dataset for Mobile Manipulation},
journal = {arXiv preprint arXiv:2509.25032},
year = {2025}
}
Release notes
- v3.0 (
v3.0tag): the same 1,184,259 episodes repacked into the LeRobot v3.0 layout - 13,318 files instead of 3.5 million, loadable with stockLeRobotDataset. Wrist-wrench padding applied to 4,608 episodes (see Limitations). - v2.1 (Storage Bucket): original per-episode layout, 3,552,783 files.
- AIRoA MoMa 5k /
airoa-moma-5k: 1,184,259 successful PA episodes, 5,025 h, five sites. - Preliminary AIRoA MoMa release: 25,469 PA episodes, approximately 94 h, one site.
- Downloads last month
- 155
Models trained or fine-tuned on airoa-org/airoa-moma-5k