CaramelCoffee19/naive-physics-ironing-v0.2

nAIve physics — Ironing Pilot v0.2 + Interaction Analysis v0.3

Visual Preview

Original RGB demonstration — IRON_009

▶ Watch IRON_009 original RGB demonstration

v0.3 interaction analysis — IRON_009

▶ Watch IRON_009 analysed interaction video

Raw → analysed: the first video is the original RGB demonstration; the second shows the v0.3 garment semantics, tool tracking, and temporal interaction analysis derived from the same episode.

A pilot real-world human garment-manipulation dataset containing RGB demonstrations of ironing / wrinkle-removal behavior.

The original v0.2 physical dataset is preserved. v0.3 is an additive interaction-analysis layer with human-verified temporal annotations, experimental analysis packages, and rendered analysis videos.

This dataset helps a model evaluate fabric slip and anchoring interactions from RGB garment-manipulation demonstrations for textile and deformable-object manipulation. The data is intended for exploratory simulation and real-robot experiments.

This release is intended for early research evaluation and feedback before collection is scaled.

Dataset Summary

• 12 manipulation episodes
• 4 physical T-shirts
• 3 target garment regions per garment
• 787.388 seconds / 13:07.388 total video
• 848 × 478 RGB
• 24 fps
• Single fixed top-down camera
• 10 cm visual spatial-reference grid
• Single operator
• 2 capture sessions

Coverage: 4 garments × 3 regions = 12 episodes.

Regions:

• front_body
• both_sleeves
• back_body

What's New in v0.3

v0.3 adds an interaction-analysis layer over the same 12 original RGB demonstrations:

• 12 human-verified temporal annotation CSVs
• 12 experimental analysis TAR packages
• 9 rendered analysis videos with garment semantics, iron/tool tracking, and temporal interaction overlays
• 77 human-verified slip events across all 12 episodes
• 1 human-verified anchor interval in IRON_006

The original v0.2 videos, metadata, calibration, garment references, and temporal annotations remain preserved.

Coverage Matrix

Physical garment	front_body	both_sleeves	back_body
GARMENT_001 — brown	Available	Available	Available
GARMENT_002 — yellow	Available	Available	Available
GARMENT_003 — pink	Available	Available	Available
GARMENT_004 — blue	Available	Available	Available

Garments

Physical garment properties are stored separately in metadata/garments.csv. The registry records color, brand, size, material composition, physical measurements, and the reference-image path for each garment.

Episode Metadata

metadata/episodes.csv contains one row per manipulation episode and references the physical garment through garment_id.

The registry includes episode ID, video filename, garment ID, capture session, operator ID, garment region, task/action, camera setup, grid spacing, duration, resolution, FPS, and bitrate.

Temporal Annotations

Original v0.2 annotations

The original human temporal annotations remain available for:

• IRON_001
• IRON_002
• IRON_004
• IRON_010

These files remain in annotations/temporal/.

Annotation source: human_annotated.

v0.3 human-verified annotations

v0.3 adds human-verified temporal annotations for all 12 episodes in:

v0.3/annotations/temporal/

The release schema is:

• episode_id
• annotation_id
• behavior
• start_time_sec
• end_time_sec
• start_frame
• end_frame
• duration_sec
• annotation_source

The v0.3 release contains:

• 77 human-verified slip events across all 12 episodes
• 1 human-verified anchor interval in IRON_006

The behavior vocabulary represented in v0.3 includes:

• initial_state
• approach
• position
• tension
• iron_stroke
• iron_hold
• reposition
• release
• terminal_state
• anchor
• slip

The v0.3 intervals originated from machine-proposed candidates but required explicit human review before inclusion. Accepted or adjusted intervals are exported as human_verified. Rejected and unreviewed candidates are excluded from the release temporal CSVs.

anchor coverage is currently sparse and should not be interpreted as comprehensive anchor annotation.

The ontology is still being evaluated during the pilot.

Data Structure

naive-physics-ironing-v0.2/
├── README.md
├── annotations/
│   └── temporal/
│       ├── IRON_001.csv
│       ├── IRON_002.csv
│       ├── IRON_004.csv
│       └── IRON_010.csv
├── calibration/
│   ├── CAL_001_grid.jpg
│   └── CAL_001_grid.mp4
├── docs/
│   ├── collection_protocol.md
│   ├── gold_episode_spec.md
│   └── schema.md
├── garment_references/
│   ├── GARMENT_001_reference.png
│   ├── GARMENT_002_reference.png
│   ├── GARMENT_003_reference.png
│   └── GARMENT_004_reference.png
├── metadata/
│   ├── episodes.csv
│   └── garments.csv
├── videos/
    ├── IRON_001_action.mp4
    ├── IRON_002_action.mp4
    ├── IRON_003_action.mp4
    ├── IRON_004_action.mp4
    ├── IRON_005_action.mp4
    ├── IRON_006_action.mp4
    ├── IRON_007_action.mp4
    ├── IRON_008_action.mp4
    ├── IRON_009_action.mp4
    ├── IRON_010_action.mp4
    ├── IRON_011_action.mp4
    └── IRON_012_action.mp4
└── v0.3/
    ├── README.md
    ├── annotations/
    │   └── temporal/
    │       ├── IRON_001.csv
    │       ├── IRON_002.csv
    │       ├── IRON_003.csv
    │       ├── IRON_004.csv
    │       ├── IRON_005.csv
    │       ├── IRON_006.csv
    │       ├── IRON_007.csv
    │       ├── IRON_008.csv
    │       ├── IRON_009.csv
    │       ├── IRON_010.csv
    │       ├── IRON_011.csv
    │       └── IRON_012.csv
    └── analysis/
        ├── packages/
        │   ├── IRON_001_experimental_analysis.tar
        │   ├── IRON_002_experimental_analysis.tar
        │   ├── IRON_003_experimental_analysis.tar
        │   ├── IRON_004_experimental_analysis.tar
        │   ├── IRON_005_experimental_analysis.tar
        │   ├── IRON_006_experimental_analysis.tar
        │   ├── IRON_007_experimental_analysis.tar
        │   ├── IRON_008_experimental_analysis.tar
        │   ├── IRON_009_experimental_analysis.tar
        │   ├── IRON_010_experimental_analysis.tar
        │   ├── IRON_011_experimental_analysis.tar
        │   └── IRON_012_experimental_analysis.tar
        └── videos/
            ├── IRON_001_analysis.webm
            ├── IRON_002_analysis.webm
            ├── IRON_003_analysis.webm
            ├── IRON_007_analysis.webm
            ├── IRON_008_analysis.webm
            ├── IRON_009_analysis.webm
            ├── IRON_010_analysis.webm
            ├── IRON_011_analysis.webm
            └── IRON_012_analysis.webm

Calibration / Spatial Reference

calibration/CAL_001_grid.jpg and calibration/CAL_001_grid.mp4 provide a 10 cm visual spatial reference.

This is NOT full intrinsic/extrinsic camera calibration.

Intended Research Use

This pilot may support exploratory research in:

• Deformable-object manipulation
• Garment manipulation
• Learning from human demonstrations
• Imitation-learning research
• Robot perception
• Physical AI
• Textile manipulation

The current pilot is not sufficient by itself to train a production robot.

Current Modalities

Available:

• Sensor-captured RGB
• Episode metadata
• Physical garment metadata
• Original human temporal annotations for 4 episodes
• Human-verified temporal annotations for 12 / 12 episodes in v0.3
• Human-verified slip annotations
• Sparse human-verified anchor annotation
• Experimental hand-landmark estimates
• Experimental garment semantics
• Approximate iron/tool tracking
• Garment-mask propagation
• Optical-flow-derived image motion
• Analysis provenance
• 9 rendered analysis videos

Not currently sensor-captured:

• Sensor-measured depth
• Thermal / IR
• Force
• Tactile
• Robot state/action vectors
• Robot proprioception
• Full camera calibration

These missing physical signals limit what robot policies and physical conclusions can be trained or evaluated from the current pilot.

Experimental Derived Data

v0.3 includes an experimental RGB-derived interaction-analysis layer.

Each episode has a package in v0.3/analysis/packages/ containing:

• semantics.csv
• hands.csv
• tool_track.csv
• slip.csv
• temporal_candidates.csv
• provenance.json

Rendered analysis videos are available for:

• IRON_001
• IRON_002
• IRON_003
• IRON_007
• IRON_008
• IRON_009
• IRON_010
• IRON_011
• IRON_012

These videos visualize available garment semantics, iron/tool tracking, and temporal interaction information.

Scientific status:

• Optical flow represents image motion, not calibrated physical velocity.
• Garment-mask propagation is approximate.
• Iron tracking is approximate.
• Model-estimated hand and garment semantics are not ground truth.
• Slip candidates require explicit human review before becoming human_verified.
• Anchor coverage is currently sparse.
• No sensor-measured depth is available.
• No force or tactile sensing is available.

Data Collection

Collection took place indoors on a fixed ironing surface using an approximately 90-degree top-down smartphone camera in a fixed position. A 10 cm grid provides a visual spatial reference, and the operator enters from camera-left.

See docs/collection_protocol.md for full details.

Dataset Files

• videos/: raw release copies of the 12 RGB episodes
• metadata/episodes.csv: episode registry
• metadata/garments.csv: physical garment registry
• annotations/temporal/: original v0.2 human temporal annotations for four episodes
• garment_references/: reference frames for the four physical garments
• calibration/: visual spatial-reference artifacts
• docs/: schema, collection protocol, and Gold Episode specification
• v0.3/annotations/temporal/: 12 human-verified temporal annotation CSVs
• v0.3/analysis/packages/: 12 experimental analysis TAR packages
• v0.3/analysis/videos/: 9 rendered analysis videos

Known Limitations

• Pilot-scale dataset
• 12 episodes
• 4 T-shirts
• Single operator
• Single environment
• Single RGB camera
• Two capture sessions
• Limited garment-type diversity
• No sensor-measured depth
• No thermal / IR
• No force/tactile
• No robot action/state data
• No formal intrinsic/extrinsic calibration
• No validated wrinkle outcome metric
• No formal train/validation/test split yet
• Derived garment semantics and iron/tool tracking are approximate
• anchor coverage is sparse
• Rendered v0.3 analysis videos currently cover 9 / 12 episodes

Outcome Measure

A validated outcome measure for whether ironing successfully removed wrinkles has not yet been defined or applied to the existing 12 episodes.

slip, anchor, temporal behavior labels, garment semantics, RGB-derived analysis, and tool tracking should therefore not be interpreted as a validated ironing-success score.

Planned Next Collection

The next planned 24 episodes are intended to expand the pilot along four axes:

Axis	Planned extension
Garment variability	Broader garment variation beyond the current four T-shirts
3D mapping	Add and evaluate reconstructed 3D representations
Thermal mapping	Add thermal / IR capture
Depth mapping	Add sensor-measured depth

These are planned extensions and are not present as sensor modalities in the current release.

Pilot Feedback

This release is intentionally public for early research evaluation.

Three researchers have provided direct feedback on the pilot so far.

Feedback is still requested on:

1. Whether human RGB garment-manipulation demonstrations are useful for their work
2. Whether the slip and anchor interaction representation is useful
3. Which missing signals or modalities matter most
4. Whether the temporal ontology is useful
5. What garment or task variation should be collected next
6. What would make this data useful for simulation or real-robot experiments

Version

Physical dataset: v0.2

Additive interaction-analysis release: v0.3

Status: public research pilot

License: CC BY-NC 4.0

The v0.3 layer does not overwrite or replace the original v0.2 release.

Links

• Website: https://naive-physics.vercel.app/
• Free Annotation Tool: https://naive-physics.vercel.app/annotate
• GitHub: https://github.com/freshlybrewedcoffee19-maker/naive-physics