---
pretty_name: wide-robot object_inside_container (real-camera v0)
task_categories:
- robotics
tags:
- robotics
- manipulation
- object-inside-container
- apriltag
- fiducial-markers
- verifier
- multi-view
size_categories:
- n<1K
license: other
license_name: tbd
---

# wide-robot — `object_inside_container` real-camera capture (v0)

Raw multi-view video for the **Phase 3A real-camera ingestion** pilot of
[wide-robot](https://github.com/alex-reysa/wide-robot): a marker-based, **source-independent
verifier** judges `object_inside_container` episodes (PASS / FAIL / UNCERTAIN) from real camera
evidence. This repo holds the **raw capture media** that is gitignored in the code repo; the derived
JSON (calibrations, tracks, rollouts, verdicts) and the full results live in the git repo under
`datasets/sony_object_inside_container_v0/` (see `INGESTION_RESULTS.md`).

> **Conservative pilot, not a benchmark.** The associated result is deliberately narrow: across 78
> clips the frozen verifier produced **0 false PASSes on 30 genuine-failure clips**, with partial
> success recall — a demonstration that the verifier's *safety* survives real, approximate-calibration
> evidence. It is marker-based and calibration-limited, not a high-recall perception system.

## Contents

- **80 clips** = 40 episodes × 2 simultaneous cameras (`sony_front` ≈ 45° frontal; `iphone_top` top
  view), under `raw_videos/<category>/oic_<category>_<seq:03d>__<camera>.mp4`.
- `manifest.json` — provenance index: per-clip SHA-256, expected class, camera, video metadata, and
  the as-built capture metadata (this is the source of truth).
- `labels.csv` — per-clip expected verdicts for the two bundled targets.
- `SHA256SUMS` — independent checksums over every `.mp4` (verify with `shasum -a 256 -c SHA256SUMS`).

### Categories
`success`, `fail_near_not_inside`, `fail_on_rim`, `fail_outside`, `control_static`,
`control_born_inside`, `control_born_inside_hand_motion`, `control_inside_to_inside`,
`control_inside_to_outside`, `success_tag_obstruction`, `success_hand_obstruction`, `calibration`.

## Capture setup (as-built; `manifest.json` is authoritative)

- Cube `5×5×5 cm`; AprilTag (`tag36h11`) IDs **2** (top) / **3** (front), 35 mm.
- Tray ~`18×18 cm`, ~4 cm rim — **homemade cardboard, not perfectly square/consistent**; ingestion
  fits the geometry. Tray tags **6** (front, outer) / **7** (inside floor), 50 mm.
- Table tags **0 / 1** (75 mm) — used only to recover the world frame. Tags 4/5 absent.

## Using it

```bash
hf download Alexreysa/wide-robot-object-inside-container-v0 --repo-type dataset --local-dir recordings
shasum -a 256 -c recordings/SHA256SUMS          # verify integrity
# then, from the wide-robot repo (pip install -e ".[camera]"):
python -m scripts.ingest_recordings --select all
```

## Provenance & licensing

- Source: tripod Sony A7IV + iPhone capture for `sony_object_inside_container_v0`; reproduced and
  judged by the byte-frozen verifier in the wide-robot repo. Every clip's SHA-256 is bound in
  `manifest.json` and `SHA256SUMS`.
- **License: TBD** — set by the dataset owner. Recommend CC-BY-4.0 (attribution) for reuse; until set,
  treat as all-rights-reserved.