airoa-org/airoa-moma

HSR Household Service Robot Teleoperation Dataset

Dataset Overview

This dataset contains 23,762 episodes of household service robot teleoperation data collected using Toyota's Human Support Robot (HSR) platform. The dataset focuses on primitive action (PA) household manipulation tasks performed through human teleoperation, providing high-quality demonstrations for robot learning research.

Key Statistics

• Total Episodes: 23,762
• Total Frames: 9,422,911 (87 hours)
• Task Success Rate: 97.4% (23,162 successful episodes)
• Average Episode Length: 396.55 frames (13.2 seconds)
• Dataset Version: 1.1
• Last Updated: November 2025
• Total Size: ~85 GB

Task Distribution

The dataset covers 7 primary household task categories:

Task Category	Episodes	Percentage	Description
Cloth Manipulation	7,573	31.9	Opening towel stand, hanging/folding towels
Coffee Making	5,581	23.5	Complete coffee preparation workflow
Dishwashing	4,470	18.9	Loading/unloading dishwasher operations
Toast Baking	3,166	13.3	Bread preparation and toaster operations
Desk Lamp Control	1,270	5.3	Button press and chain pull light control
Slipper Organization	1,006	4.2	Arranging slippers in rack
Other Tasks	696	2.9	Miscellaneous household tasks

Complete Short Horizon Task List

The dataset contains 7 distinct short horizon tasks that combine multiple primitive actions:

Rank	Short Horizon Task	Episodes	Percentage	Task Category
1	Open the towel stand and hang the towel.	7,573	31.9	Cloth Manipulation
2	Make coffee	5,581	23.5	Coffee Making
3	Washing dishes in the dishwasher	4,470	18.8	Dishwashing
4	Bake a toast	3,166	13.3	Toast Baking
5	Press the button to turn the desk lamp on and off	1,270	5.3	Desk Lamp Control
6	Stand the slippers in the slipper rack	1,006	4.2	Slipper Organization
7	Pull the chain to turn the desk lamp on or off	696	2.9	Desk Lamp Control

Top Individual Primitive Tasks

1. Open the towel stand. (1,297 episodes)
2. Put the towel into the basket. (1,295 episodes)
3. Grab the towel hanging on the towel stand. (1,292 episodes)
4. Hang the towel on the towel stand. (1,276 episodes)
5. Grab the towel in the basket. (1,218 episodes)

Basic Skill Distribution

Analysis of primitive actions reveals the fundamental manipulation skills required:

Rank	Basic Skill	Occurrences	Percentage	Description
1	Grab	3252	13.7%	Grasping and holding objects
2	Place	3197	13.5%	Placing objects in specific locations
3	Open	3182	13.4%	Opening containers, doors, stands
4	Close	1897	8.0%	Closing containers, doors, lids
5	Put	1852	7.8%	Putting objects into containers
6	Press	1542	6.5%	Pressing buttons and controls
7	Pick	1525	6.4%	Picking up objects from surfaces
8	Pull	1464	6.2%	Pulling chains, handles, drawers
9	Hang	1276	5.4%	Hanging objects on stands/hooks
10	Fold	1195	5.0%	Folding towel stands and objects
11	Push	1010	4.3%	Pushing buttons, trays, objects
12	Take	747	3.1%	Taking objects from locations
13	Approach	426	1.8%	Moving toward target locations
14	Move	421	1.8%	Moving away from target locations
15	Insert	343	1.4%	Inserting objects into slots
16	Remove	226	1.0%	Removing objects from containers
17	Run	207	0.9%	Run the target objects

Total Primitive Actions: 23,762 across all episodes

The distribution shows a balanced representation of fundamental manipulation skills, with emphasis on:

• Container manipulation (Open/Close): 21.4% of all actions
• Object placement (Place/Put): 21.2% of all actions
• Object acquisition (Grab/Pick/Take): 23.2% of all actions
• Force application (Pull/Push/Press): 16.9% of all actions

Data Structure

Video Data

• Camera Setup: Dual RGB cameras (hand-mounted + head-mounted)
• Resolution: 640×480 pixels
• Framerate: 30 frames per second
• Format: MP4 files
• Camera Calibration: Included for both cameras with distortion parameters

Metadata (episodes.jsonl)

Each episode contains comprehensive metadata:

{
  "episode_index": 0,
  "tasks": ["Pull the chain to turn off the light."],
  "length": 729,
  "uuid": "a699601f-41e5-4678-865d-d9de37a010ad",
  "task_type": "PA",
  "task_success": true,
  "short_horizon_task": "Pull the chain to turn the desk lamp on or off",
  "primitive_action": ["Action sequence"],
  "label": "Operator001",
  "hsr_id": "robot003",
  "location_name": "location001",
  "calib": {...},
  "version": "1.0",
  "git_hash": "v4.0.0"
}

Key Metadata Fields

• episode_index: Unique episode identifier (0-25468)
• tasks: List of specific tasks performed in episode
• length: Number of timesteps in episode
• task_success: Boolean indicating task completion success
• short_horizon_task: High-level task description
• primitive_action: Detailed action sequence breakdown
• calib: Camera calibration parameters for head and hand cameras
• uuid: Uniquie identifier of high-level episode.
• label: Anonimized operator identifier.

Hardware Configuration

Robots

• Platform: Toyota Human Support Robot (HSR)
• Count: 8 robots (anonymized as robot001-robot008)
• Distribution:
  • robot002: 8,685 episodes (34.1%)
  • robot001: 4,286 episodes (16.8%)
  • robot005: 4,132 episodes (16.2%)
  • robot004: 3,068 episodes (12.0%)
  • Others: <10% each

Human Operators

• Count: 19 teleoperators (anonymized as Operator001-Operator019)
• Interface: HSR leader teleoperation system
• Primary Contributors:
  • Operator015: 12,408 episodes (48.7%)
  • Operator009: 3,071 episodes (12.1%)
  • Operator003: 2,175 episodes (8.5%)

Environment

• Location: Single laboratory environment (anonymized as location001)
• Setup: Controlled household environment with kitchen appliances and furniture

Data Anonymization

All personally identifiable information has been systematically anonymized using the mapping system defined in anonymization_mappings.json:

Anonymization Mappings

• Human Operators: 19 operators → Operator001-Operator019
• Robot IDs: 8 HSR units → robot001-robot008
• Locations: 1 lab environment → location001
• Git Hashes: Development commits → semantic versions (v1.0.0-v12.0.0)

Technical Specifications

Camera Calibration

Both cameras include complete calibration parameters:

• Intrinsic Matrix (K): 3×3 camera matrix
• Distortion Coefficients (D): Radial and tangential distortion
• Projection Matrix (P): 3×4 projection matrix
• Rectification Matrix (R): 3×3 rectification matrix

Usage Guidelines

Research Applications

• Robot Learning: Imitation learning from teleoperation demonstrations
• Computer Vision: Multi-view manipulation task understanding
• Task Planning: Hierarchical task decomposition analysis
• Human-Robot Interaction: Teleoperation interface studies

Quality Metrics

• Task Success Rate: 97.4% overall success rate
• Episode Length Distribution: 120-856 frames (avg: 399.73)
• Data Completeness: All episodes have corresponding hand and head camera videos
• Annotation Quality: Rich task decomposition with primitive action sequences

Limitations

• Environment Scope: Single laboratory setting may limit generalization
• Task Diversity: Focus on specific household tasks (7 main categories)
• Operator Variance: Uneven distribution across human operators
• Temporal Scope: Data collected during specific development phases

How to Download

Since the dataset have a lot of files, Hugging Face API can hit rate limit easily.

$ hf download airoa-org/airoa-moma --repo-type dataset

...

We had to rate limit you, you hit the quota of 1000 api requests per 5 minutes period. Upgrade to a PRO user or Team/Enterprise organization account (https://hf.co/pricing) to get higher limits. See https://huggingface.co/docs/hub/rate-limits

So we recommend you to download with Git over SSH.

First, follow the official instruction and set your public SSH key to Hugging Face.

cd ~/.ssh
ssh-keygen -t ed25519 -C "[email protected]" -f <key file>

Second, install Git LFS, which manages large files like videos. After installing Git LFS, you need to execute git lfs install once per user.

Then set up SSH Agent otherwise you will be asked SSH key pass phrase per files. One of the good documents is an istruction at GitHub Docs.

eval "$(ssh-agent -s)"
ssh-add ~/.ssh/<key file>

Additionally, add an entry at ~/.ssh/config

Host hf.co
  User git
  IdentityFile ~/.ssh/<key file>

Finally, you can clone;

git clone [email protected]:datasets/airoa-org/airoa-moma.git
cd airoa-moma
git lfs pull

Depending on your git configuration, git lfs pull can be run during git clone automatically. If you haven't pulled LFS managed files, they are just pointer text files.

$ file videos/chunk-000/observation.image.hand/episode_000000.mp4
videos/chunk-000/observation.image.hand/episode_000000.mp4: ASCII text

$ cat videos/chunk-000/observation.image.hand/episode_000000.mp4
version https://git-lfs.github.com/spec/v1
oid sha256:48277551133b1587c4c02cec6ee41f9a925565cf4d8aa9d0931f1d997d39c0a6
size 8488109

Once you pull LFS files, then they become regular files.

$ file videos/chunk-000/observation.image.hand/episode_000000.mp4
videos/chunk-000/observation.image.hand/episode_000000.mp4: ISO Media, MP4 Base Media v1 [ISO 14496-12:2003]

Change Log

• v1.1: Filter out skeptical episodes
  • Too short (< 1.0s) or too long (> 60.s) episodes
  • Large jump (> 1.0) at any dimensions of observation.state
  • Delay longer than 0.3s at any of observation.state, observation.image.head, or observation.image.hand

Citation

If you use this dataset in your research, please cite:

@article{airoa-moma-2025,
  author  = {Ryosuke Takanami, Petr Khrapchenkov, Shu Morikuni, Jumpei Arima, Yuta Takaba, Shunsuke Maeda, Takuya Okubo, Genki Sano, Satoshi Sekioka, Aoi Kadoya, Motonari Kambara, Naoya Nishiura, Haruto Suzuki, Takanori Yoshimoto, Koya Sakamoto, Shinnosuke Ono, Yo Ko, Daichi Yashima, Aoi Horo, Tomohiro Motoda, Kensuke Chiyoma, Hiroshi Ito, Koki Fukuda, Akihito Goto, Kazumi Morinaga, Yuya Ikeda, Riko Kawada, Masaki Yoshikawa, Norio Kosuge, Yuki Noguchi, Kei Ota, Tatsuya Matsushima, Yusuke Iwasawa, Yutaka Matsuo, Tetsuya Ogata},
  title   = {AIRoA MoMa Dataset: A Large-Scale Hierarchical Dataset for Mobile Manipulation},
  journal = {arXiv preprint},
  year    = {2025}
}