FieldData Outdoor Labor Motion Dataset (FOLMD) is a multi-modal motion dataset of professional outdoor workers performing real-world manual labor tasks in urban environments in Japan.

This dataset is designed to support the development of physical AI foundation models and embodied robotic systems targeting outdoor maintenance and construction tasks. All data was collected from professional workers at actual job sites — not from staged laboratory environments.

The v1.0 release covers a complete two-phase urban fence restoration task:

Graffiti removal using chemical solvent and rag wiping
Repainting with black paint using brushes at multiple heights and postures

Data modalities include egocentric RGB video, depth maps (Intel RealSense D435i), full-body IMU (8-point, Mbientlab MetaMotionS), glove-based grasp force (Tekscan Grip System), and annotated 3D skeleton sequences (SMPL format).

Why This Dataset?

Challenge for Robots	What This Dataset Provides
Vertical surface contact control	Sustained brush pressure on metal slat fence (9–25 N)
Low-posture precision work	Deep crouch / kneeling brushwork at ground level
Boundary-aware manipulation	Painting within 10 mm of masking tape edges
Multi-posture task transitions	Standing → crouching → kneeling within a single session
Visual quality assessment	Workers pausing to inspect coverage, detecting missed spots
Multi-agent coordination	3–4 workers dividing fence height and working in parallel

These tasks involve sustained contact with irregular vertical surfaces, fine motor control under physical load, and situated decision-making — all critical challenges for next-generation manipulation models.

Supported Tasks

Imitation Learning (IL): Action-annotated demonstrations suitable for behavior cloning and inverse reinforcement learning
Vision-Language-Action (VLA): Paired egocentric video and action labels for multimodal training
Pose Estimation: Full-body SMPL skeleton data across 6 posture categories
Action Segmentation: 7-class taxonomy with fine-grained sub-actions and temporal boundaries
Force Control Learning: Grasp pressure profiles for 5 grip types across different tools and surfaces
Contact-Rich Manipulation: Depth + force data for learning compliant surface following

Data Collection

Sensor Setup

Modality	Device	Specs	Sample Rate
Egocentric RGB	GoPro Hero 12	1920×1080	60 fps
Depth + RGB	Intel RealSense D435i	848×480	30 fps
Full-body IMU (×8)	Mbientlab MetaMotionS	Accel + Gyro + Euler	200 Hz
Grasp Force Glove	Tekscan Grip System	24 cells, right hand	100 Hz
3D Skeleton	Computed from IMU	SMPL format	30 fps
External cameras	GoPro Hero 12 (×3)	1920×1080	60 fps

Temporal synchronization: All sensors synchronized via GPS-PPS hardware sync signal. Timestamp accuracy: ±0.8 ms.

Sensor Placement

Head:     GoPro (front) + RealSense (side)
R Wrist:  IMU
L Wrist:  IMU
R Elbow:  IMU
L Elbow:  IMU
Waist:    IMU
R Knee:   IMU
L Knee:   IMU
R Hand:   Force glove (Tekscan)

Collection Protocol

Workers are professional outdoor maintenance staff with 2–10 years of experience
Written informed consent obtained from all participants
Data collected at real urban job sites (Tokyo metropolitan area, Japan)
Session begins with 5-minute sensor calibration and zeroing
Workers perform tasks naturally without scripted movements
All sessions include workers of mixed skill levels (novice / intermediate / experienced)

Dataset Structure

folmd-v1.0/
├── sessions/
│   └── session_20210327_001/
│       ├── raw/
│       │   ├── gopro_head.mp4            # Egocentric RGB (60 fps)
│       │   ├── realsense_rgb.mp4         # RealSense RGB (30 fps)
│       │   ├── realsense_depth.bag       # Depth stream (ROS bag)
│       │   ├── imu_r_wrist.csv           # Right wrist IMU
│       │   ├── imu_l_wrist.csv           # Left wrist IMU
│       │   ├── imu_r_elbow.csv           # Right elbow IMU
│       │   ├── imu_l_elbow.csv           # Left elbow IMU
│       │   ├── imu_waist.csv             # Waist IMU
│       │   ├── imu_r_knee.csv            # Right knee IMU
│       │   ├── imu_l_knee.csv            # Left knee IMU
│       │   ├── imu_head.csv              # Head IMU
│       │   └── force_glove_right.csv     # Right hand force (24 cells)
│       ├── processed/
│       │   ├── skeleton_smpl/
│       │   │   └── frame_*.json          # Per-keyframe SMPL skeleton
│       │   ├── depth_pointcloud/
│       │   │   └── frame_*.pcd           # Per-keyframe 3D point cloud
│       │   └── action_segments.json      # Full annotation (Tier 2)
│       └── metadata/
│           ├── session_info.json
│           ├── sensor_calibration.json
│           └── quality_report.pdf
├── dataset_card.md                       # This file
├── annotation_guidelines.pdf            # Labeling rulebook (EN)
└── sample/
    └── 30min_annotated_sample/           # Free sample — see Licensing

Annotations

Action Taxonomy (7 Primary Labels)

Action Label	Description	Avg Duration (sec)
`solvent_application`	Applying chemical solvent to dissolve graffiti using rag	45–180
`scrub_with_rag`	Physical scrubbing to remove dissolved paint	10–60
`paint_application`	Applying paint with brush/roller at mid-to-upper height	30–300
`detail_paint_application`	Precision brushwork at edges, slat gaps, base boundaries	10–120
`visual_inspection`	Standing back to scan surface for quality assessment	5–30
`material_preparation`	Handling tools, mixing paint, preparing supplies	5–60
`masking`	Applying or adjusting masking tape	30–180

Posture Labels (6 Categories)

standing_upright          — full height, minimal trunk lean
standing_slight_lean      — 10–20° trunk forward flexion
crouching_moderate        — knee flexion 60–90°
crouching_deep            — knee flexion > 90°
kneeling_single_knee      — one knee on ground
kneeling_double_knee      — both knees on ground

Force Labels (Qualitative)

minimal    < 5 N    (passive hold / inspection)
low        5–10 N   (precision brush / fine detail)
medium     10–20 N  (normal painting stroke)
high       20–30 N  (scrubbing / removal force)
very_high  > 30 N   (heavy scrub on resistant surface)

Annotation Layers

Layer	Content	Completeness
Layer 1	Action segmentation (start/end time + label)	100%
Layer 2	Object + surface + environment labels	100%
Layer 3	Skill quality label (expert/intermediate/novice)	100%
Layer 4	Failure cases and recovery actions	100%

Inter-annotator agreement (Cohen's κ): 0.94

Data Fields

`action_segments.json`

Field	Type	Description
`session_id`	string	Unique session identifier
`frame_id`	string	Keyframe identifier
`timestamp_sec`	float	Time from session start (seconds)
`phase`	string	`graffiti_removal` or `repaint_application`
`action_label`	string	Primary action (see taxonomy above)
`sub_action`	string	Fine-grained sub-action description
`worker_id`	string	Worker identifier (W01–W04)
`body_posture`	string	Posture category
`working_height`	string	Vertical zone on fence being worked

`imu_*.csv`

Field	Type	Unit	Description
`timestamp_sec`	float	s	Synchronized timestamp
`accel_x/y/z_g`	float	g	Linear acceleration
`gyro_x/y/z_dps`	float	°/s	Angular velocity
`euler_roll/pitch/yaw_deg`	float	°	Fused orientation estimate

`force_glove_right.csv`

Field	Type	Unit	Description
`timestamp_sec`	float	s	Synchronized timestamp
`cell_{name}_N`	float	N	Force at each of 24 sensor cells
`total_force_N`	float	N	Sum across all active cells
`grip_type`	string	—	Inferred grip classification

`realsense_depth.bag` (ROS bag)

Topic	Format	Description
`/camera/depth/image_rect_raw`	sensor_msgs/Image	16-bit depth (mm)
`/camera/color/image_raw`	sensor_msgs/Image	RGB aligned to depth
`/camera/depth/color/points`	sensor_msgs/PointCloud2	Fused RGBD point cloud

Data Splits

Split	Sessions	Hours	Notes
Sample (free)	1	0.5	Available without license agreement
Full v1.0	10	50	Commercial license required
Planned v2.0	50+	300+	Additional tasks: garbage collection, weeding, building cleaning

Dataset Creation

Motivation

Physical AI and embodied robot models are rapidly advancing, but publicly available training data for outdoor real-world manipulation tasks remains extremely scarce. Existing motion datasets focus primarily on laboratory manipulation, household tasks, or driving — leaving a large gap in construction, maintenance, and urban service work.

This dataset addresses that gap by collecting data from professional Japanese workers performing skilled outdoor maintenance tasks. Japan's workforce is recognized globally for precision, safety discipline, and consistent execution — making it an ideal source of high-quality demonstration data.

Collection Methodology

Task protocols defined in cooperation with professional site supervisors
Sessions conducted at real operational job sites
Multiple skill levels included to capture expert vs. novice motion differences
Both "successful" and "recovery from failure" motion captured and labeled
All data cleaned, synchronized, and formatted to be pipeline-ready

Known Limitations

All data collected in Tokyo metropolitan area — urban environment bias
Nighttime and rainy-weather sessions represent < 15% of current dataset
Force glove covers right hand only
Skeleton estimation accuracy degrades at extreme joint angles (> 110°)

Licensing Information

This dataset is released under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

✅ Free to use for academic and non-commercial research
✅ Free 30-minute sample available without registration
✅ Redistribution permitted with attribution
❌ Commercial use requires a separate commercial license

For commercial licensing inquiries, please contact: contact@fielddata.jp

Citation

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

@dataset{fielddata2025folmd,
  author    = {FieldData Japan},
  title     = {FieldData Outdoor Labor Motion Dataset (FOLMD) v1.0},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/fielddata-jp/folmd},
  note      = {Multi-modal motion dataset for physical AI training — outdoor maintenance tasks, Tokyo, Japan}
}