# HANDFORGE ROBOTICS — DATA CARD

Delivery specification for licensed manipulation corpora and commissioned capture
programmes. Version 3 taxonomy.

Handforge Robotics is in formation. Capture programmes open in 2027. This card
describes what a delivered drop contains and the terms it arrives under, so that
ingestion and legal review can happen before anything is recorded.

---

## 1. WHAT AN EPISODE IS

One attempt at one task, by one named operator, in one instrumented environment.

An episode is not a clip. It is bounded by the task's start and end conditions,
scored by a person against a written acceptance criterion, and shipped with the
provenance that lets a buyer defend the purchase. Failed attempts are retained and
labelled rather than discarded — a corpus of only successes cannot teach recovery.

## 2. MODALITIES

| Channel | Specification |
|---|---|
| Visual | 4 × RGB-D viewpoints, 30 fps, hardware-synchronised. Episodes exceeding 2 ms measured drift are rejected at QA. |
| Depth | Stereo pair plus active depth. Per-frame point cloud retained, not derived on request. |
| Hand | 6-DoF wrist pose and 21-point skeleton at 60 Hz, expressed in the task frame. |
| Contact | Per-fingertip normal force where instrumented gloves were fitted. Flagged per episode; not universal. |
| Objects | Per-frame masks and 6-DoF pose for task-relevant objects only. Background is not labelled. |
| Audio | Not captured. Guest and worker environments make it a privacy liability without a training case. |

## 3. WHAT TRANSFERS TO YOUR EMBODIMENT

Delivered in three layers, so you retarget with your own method rather than inherit ours.

**Object layer — embodiment-independent.** Object masks and 6-DoF pose over time.
The state change is the task; it carries no morphology.

**Hand layer — retargetable.** Wrist pose plus full skeleton. For a parallel gripper,
aperture derives from thumb-to-index distance and approach from the wrist frame; for a
multi-finger hand, the skeleton retargets joint-to-joint. Both the derived end-effector
trajectory and the raw skeleton ship, so nothing is lost to our choice of mapping.

**Outcome layer.** Pass or fail plus failure class, so the corpus can be filtered and
weighted by result.

A motion gap survives retargeting and is closed with teleoperation capture on your
own hardware. We do not claim human demonstration removes the need for on-robot data.

## 4. DROP STRUCTURE

```
handforge_<class>_v<n>/
  manifest.json         drop id, episode counts, per-file checksums
  DATA_CARD.md          this document, pinned to the drop
  episodes/<id>/        episode.json, cam_00..03.mp4, depth.parquet,
                        hand.parquet, objects.parquet
  index/episodes.parquet  queryable index across the drop
  index/lerobot/          LeRobot conversion
  index/rlds/             RLDS shards
```

Schema for `episode.json`: `handforge-episode-schema.json` (JSON Schema 2020-12).
Formats: MP4 media, Parquet sidecars, LeRobot and RLDS conversions. HDF5 on request.

## 5. PROVENANCE AND CONSENT

Every episode names the operator by stable pseudonymous identifier, with role and
years in role attached. Legal names never appear in delivered data.

- Participation is consented in writing, separately from employment, and is revocable.
- Capture time is paid at shift rates. No episode is recorded off the clock.
- Faces are excluded from delivered data by default.
- A revoked consent triggers removal from subsequent drops and a deletion certificate
  for episodes already delivered.

## 6. WHAT IS NOT IN A LICENSED DROP

- **Held-out environments.** A subset of sites is permanently ring-fenced for
  evaluation and never appears in any corpus offered for sale.
- **Another customer's commissioned data.** Programme output is never resold into a
  general corpus during an exclusivity window.
- **Synthetic or simulated episodes.** Everything in a drop was performed by a person
  and recorded at the time it happened.
- **Scraped or third-party video.**

## 7. LICENCE

| Term | Detail |
|---|---|
| Default | Non-exclusive, perpetual, worldwide, for training and evaluating models. |
| Exclusive | 12, 18 or 24-month windows on commissioned programmes, reverting to non-exclusive. |
| Redistribution | Not permitted. Model weights trained on the corpus are yours without restriction. |
| Evaluation licence | Sample drops only. Evaluation and integration testing; not for production training. |
| Deletion | Contractual deletion certificates on termination or consent withdrawal. |

## 8. JURISDICTION

GDPR processor terms and EU AI Act training, validation and test-data governance for
EU buyers; CBPR-aligned cross-border terms for Japan; China-local processing with no
offshore raw data; Philippine Data Privacy Act 2012 and National Privacy Commission
obligations at source. Data processing addendum, subprocessor list and retention
schedule available before contracting.

## 9. KNOWN LIMITATIONS

Stated plainly, because a buyer finds them anyway.

- Force data is not universal. It is present only where the task warranted gloves.
- Environment diversity is concentrated in Metro Manila. Lighting, fixtures and linen
  types reflect Philippine hospitality and fitness estates.
- Object pose accuracy degrades under heavy occlusion, which is common in confined
  spaces. Affected frames are flagged rather than interpolated.
- The taxonomy is versioned and will change. Drops are pinned to the version they were
  captured under; taxonomy migrations are documented, not silent.
- Human demonstration reduces but does not eliminate the on-robot data requirement.

---

Questions on any of the above, before commercial discussion: hello@handforgerobotics.com