README.md

---
license: cc-by-4.0
language:
- en
pretty_name: ARMBench Defect Detection - Video Subset
tags:
- armbench
- amazon-robotics
- robotic-manipulation
- defect-detection
- video-classification
- warehouse-automation
task_categories:
- video-classification
size_categories:
- 100K<n<1M
annotations_creators:
- expert-generated
source_datasets:
- original
configs:
- config_name: labels
  default: true
  data_files:
  - split: train
    path: labels/train-*.parquet
  - split: test
    path: labels/test-*.parquet
---

# ARMBench - Defect Detection, Video Subset

This repository is the official Amazon redistribution, on the Hugging Face Hub, of the **video defect
detection** subset of ARMBench. The media payload is Amazon's original release archive, byte-for-byte,
with its published SHA-256 intact; the small label, annotation and manifest files alongside it are
derived artifacts added so that the labels can be read without downloading 254 GiB.

Project home: <http://armbench.com> &nbsp;|&nbsp; Paper: [arXiv:2303.16382](https://arxiv.org/abs/2303.16382)
&nbsp;|&nbsp; Contact: armbench-support@amazon.com

> **Note on the project website.** `http://armbench.com` is Amazon's canonical home for ARMBench and is
> the URL printed in the paper. At the time of writing the `https://www.armbench.com` certificate has
> expired, so some browsers will warn before showing the site. This has been reported to the ARMBench
> team. The arXiv link above is a non-expiring alternative reference.

## Dataset Description

The following description is quoted verbatim from Amazon's ARMBench website.

> **What is ARMBench?**
>
> ARMBench is a large-scale benchmark dataset for perception and manipulation challenges across Amazon
> Robotics systems. The datasets are collected in Amazon warehouses and capture a wide variety of objects
> and configurations. They comprise images, videos, and tabular data for different stages of robotic
> manipulation with high-quality annotations.

ARMBench currently includes datasets from two Amazon Robotics programs. This subset belongs to the
**Sparrow Tasks** group, described on the site as:

> Sparrow - Amazon's intelligent robotic system that streamlines the fulfillment process by moving
> individual products before they get packaged. Sparrow is a robotic system in Amazon warehouses that can
> detect, select, and handle individual products in inventory using computer vision and artificial
> intelligence. It focuses on object segmentation, identification, and defect detection, enabling more
> efficient and safer operations while allowing employees to focus on higher-value tasks.

ARMBench documents **five benchmark tasks** and distributes them as **six downloads**: Defect Detection
ships as two separate downloads (images and videos), and the three Object Segmentation subsets ship
inside a single download. This repository is one of those six downloads; the others are published under
the same [`amazon`](https://huggingface.co/amazon) organization with the `armbench-` prefix.

### The defect detection task, verbatim

> The defect detection task is to identify if a robotic manipulation activity resulted in a defect. Two
> types of robot-induced defects are included in the dataset: 1. multi-pick, and 2. package-defect.
> Multi-pick is used to describe activities where multiple objects were picked and transferred from the
> source container to the destination container. Package-defect is used to describe activities where the
> object packaging *opened* and/or the object separated into multiple parts (*deconstruction*). Two
> subclasses, *open* and *deconstruction*, are defined for package-defect.

### This subset, verbatim

> **Video defect detection (255G):** This dataset comprises 4,075 videos of objects with defects.
> Multi-picks are not observable in videos and are excluded from this dataset. At the same time, open and
> deconstruction defects are observable in videos and are annotated. 100,000 videos of activities that did
> not result in a defect are available in the dataset.
>
> **Train-Test-Split:** List of training and test cases.

Two editorial notes on the quotations above, so that the differences from the live page are auditable and
not mistaken for transcription errors. First, parentheses and italics were added to
`...separated into multiple parts (*deconstruction*)`; the live page italicises the term but omits the
parentheses. Second, the live page's sentence "For image defection detection, ..." belongs to the image
subset and is not quoted here. Nothing else was changed, added or reordered.

## Dataset Structure

### What this repository contains

| Path | Bytes | Description |
|---|---|---|
| `archive/armbench-defects-video-0.1.tar.gz` | 272,781,053,674 | Amazon's original release archive, unmodified |
| `archive/armbench-defects-video-0.1.tar.gz.sha256` | 102 | Amazon's published checksum sidecar, unmodified |
| `archive/armbench-defects-tiny.tar.gz` | 69,609,824 | Amazon's small defect-detection sample archive (covers both the image and video subsets), unmodified |
| `labels/train-00000-of-00001.parquet` | - | 8,757 rows, derived (see schema below) |
| `labels/test-00000-of-00001.parquet` | - | 11,251 rows, derived |
| `annotations/train.csv` | 192,654 | Amazon's official train split list, unmodified |
| `annotations/test.csv` | 247,522 | Amazon's official test split list, unmodified |
| `annotations/video_annotations_104075.jsonl` | 8,649,828 | All 104,075 per-video annotation JSONs, one per line, derived |
| `annotations/video_defect_labels_ALL_104075.csv` | 5,068,525 | Flattened clip-level labels for all 104,075 videos, derived |
| `annotations/archive_listing_verbose.txt` | 19,774,629 | `tar -tvzf` listing of every archive member with its size, 208,156 lines, derived |
| `baseline/defect_detection_videos.ckpt` | 436,120,692 | Amazon's released baseline checkpoint, unmodified |
| `License-CC-BY.txt` | - | The licence file shipped at the root of Amazon's archive, unmodified |

### Archive contents

The archive holds, at its root:

- `videos/<pick_id>.mp4` - 104,075 videos, 1280x560
- `annotations/<pick_id>.json` - 104,075 annotation files, one per video, 1:1 with the videos
- `train.csv`, `test.csv` - the official split lists
- `License-CC-BY.txt`

**The annotation files sit at the end of the gzip stream.** A partial download, or `tar tzf ... | head`,
shows only `videos/` and the archive will appear to be unlabelled. It is not. The gzip stream is not
seekable, so any full listing or extraction is a single sequential pass (roughly 25 minutes on a modern
machine). If you only need labels, use `labels/` or `annotations/` in this repository and never touch the
archive.

### `labels` configuration (the default, and what the viewer shows)

One row per pick in the official train/test split. 20,008 rows total.

| Column | Type | Description |
|---|---|---|
| `pick_id` | string | Pick identifier; the archive member is `videos/<pick_id>.mp4` |
| `split` | string | `train` or `test`, from Amazon's own `train.csv` / `test.csv` |
| `rel_path` | string | Path of the video inside the archive |
| `label` | string | Clip-level class: `nominal`, `open` or `deconstruction` |
| `y_open` | int8 | Multi-hot target channel 0 |
| `y_deconstruction` | int8 | Multi-hot target channel 1 |
| `ev_open` | int8 | 1 if any frame-level event in this clip is an `open` event |
| `ev_deconstruction` | int8 | 1 if any frame-level event in this clip is a `deconstruction` event |
| `sublabel` | string | Item type for defect clips: `book`, `box`, `bag`, `other`; empty for nominal |
| `defect_onset_frames` | string | Semicolon-separated frame indices of the events matching the clip label; empty for nominal |
| `onset_frames_all_events` | string | Semicolon-separated frame indices of **all** annotated events, including events whose class differs from the clip label |

The raw per-video annotation schema, as shipped inside the archive and reproduced line-by-line in
`annotations/video_annotations_104075.jsonl`, is
`{"id", "label", "sublabel", "annotations": [{"frame_idx", "label"}]}`.

### Splits

| Split | Videos | nominal | deconstruction | open | Defect rate |
|---|---|---|---|---|---|
| train | 8,757 | 5,838 | 1,520 | 1,399 | 33.3% |
| test | 11,251 | 10,000 | 645 | 606 | 11.1% |
| **total** | **20,008** | **15,838** | **2,165** | **2,005** | 20.8% |

**The test split is deliberately larger than the train split.** Train is defect-enriched (33.3% defects)
to make learning tractable; test approximates the operational rate (11.1%). This is by design. Do not
rebalance it if you are reporting benchmark numbers.

The remaining **84,067 videos** in the archive (205.485 GiB, 80.60% of the archive's video bytes) are not
part of the official split. They carry annotations and can be used for pre-training or self-supervision,
but they are not part of the benchmark. Clip-level labels for all 104,075 videos, in and out of split, are
in `annotations/video_defect_labels_ALL_104075.csv`.

Sizes of the useful subsets, measured from the archive's own member table:

| Tier | Videos | Bytes | Size |
|---|---|---|---|
| Whole archive | 104,075 | 272,781,053,674 | 254.047 GiB |
| Official train + test | 20,008 | 53,094,867,799 | 49.448 GiB |
| Official test only | 11,251 | 29,747,682,304 | 27.705 GiB |
| Defect videos only | 4,170 | - | 10.849 GiB |
| Labels and annotations only | - | ~34,000,000 | ~33 MiB |

### The benchmark target

The official target is **multi-hot over two independent sigmoid channels**, not a 3-way softmax and not a
binary defect/no-defect collapse:

| Clip label | `[y_open, y_deconstruction]` |
|---|---|
| `nominal` | `[0, 0]` |
| `open` | `[1, 0]` |
| `deconstruction` | `[0, 1]` |

Amazon's released baseline, `baseline/defect_detection_videos.ckpt`, has a `(2, 768)` classification head
trained with binary cross-entropy, which is the reason to keep two channels rather than three classes.

At the **clip** level the pair `[1, 1]` does not occur in this release, so a 3-class model would be
numerically equivalent on these labels. At the **frame-event** level the two classes do co-occur: 946 of
the 20,008 clips contain both an `open` event and a `deconstruction` event even though their clip label
names only one. Models that consume frame events need the two channels.

## Known data quirks

These are measured properties of the release, documented so you do not have to rediscover them.

1. **Frame-event label strings carry trailing spaces.** The event vocabulary includes `"Open "` and
   `"Deconstruction "` with a trailing space, as well as `"No Damage"` and null. Strip before comparing.
2. **Some `frame_idx` values exceed the true frame count of the corresponding mp4.** Clamp to the last
   frame; do not subtract an offset.
3. **`defect_onset_frames` is a strict subset of `onset_frames_all_events` in 343 of 20,008 rows**
   (1.7%). If you need every annotated event, read `onset_frames_all_events`.
4. **15,838 rows have an empty `defect_onset_frames`.** These are exactly the nominal clips.
5. **gzip buys almost nothing here** - 0.416% on already-encoded mp4. The archive is gzipped because
   Amazon released it that way, not because it compresses.
6. `annotations/archive_listing_verbose.txt` is Amazon's member listing with sizes. In some earlier
   working copies this file was misleadingly named `partial_tar_listing.txt`; it is the **complete**
   208,156-line verbose listing, not a partial one.

## Published figures versus measured figures

Amazon's published figures are given first; the figures measured from this exact release follow, labelled
as such. Both are reported rather than silently reconciled.

| Quantity | Amazon's published figure | Measured from this release |
|---|---|---|
| Defect videos | 4,075 (task page), 4,000+ (home page) | **4,170** (2,165 deconstruction + 2,005 open), over all 104,075 annotations |
| Non-defect videos | 100,000 (task page), 100,000+ (home page) | **99,905** nominal (95.99% of 104,075) |
| Download size | 255G (task page), 254 GB (release email) | **272,781,053,674 bytes** = 254.047 GiB = 272.78 GB |

The 4,075 / 4,170 gap has been raised with the ARMBench team. Until it is resolved, treat 4,170 as the
count you will actually find in the files and 4,075 as the published figure.

## Verifying what you downloaded

The archive in this repository is byte-identical to Amazon's release object, so Amazon's own published
checksum verifies it:

```bash
hf download amazon/armbench-defect-video \
  archive/armbench-defects-video-0.1.tar.gz \
  archive/armbench-defects-video-0.1.tar.gz.sha256 \
  --repo-type dataset --local-dir ./armbench-defect-video

cd ./armbench-defect-video/archive
sha256sum -c armbench-defects-video-0.1.tar.gz.sha256
# expected: a0dccbc3e92f7587bf9bd8fa5f57facb7ce7e04b4303680847dbdbe62aaa9905
```

Labels only, no archive:

```python
from datasets import load_dataset
ds = load_dataset("amazon/armbench-defect-video")   # the `labels` config, 20,008 rows
print(ds["train"][0])
```

## Dataset Creation

The data was collected in Amazon warehouses from a robotic manipulator performing object singulation from
containers with heterogeneous contents, as described in the ARMBench paper:

> This paper introduces Amazon Robotic Manipulation Benchmark (ARMBench), a large-scale, object-centric
> benchmark dataset for robotic manipulation in the context of a warehouse. [...] We present a
> large-scale dataset collected in an Amazon warehouse using a robotic manipulator performing object
> singulation from containers with heterogeneous contents. ARMBench contains images, videos, and metadata
> that corresponds to 235K+ pick-and-place activities on 190K+ unique objects. The data is captured at
> different stages of manipulation, i.e., pre-pick, during transfer, and after placement. Benchmark tasks
> are proposed by virtue of high-quality annotations and baseline performance evaluation are presented on
> three visual perception challenges, namely 1) object segmentation in clutter, 2) object identification,
> and 3) defect detection. ARMBench can be accessed at http://armbench.com

Clip-level and frame-level defect annotations were produced by human annotators. Amazon's website does not
publish the annotation protocol, the annotator population, or inter-annotator agreement, and this
redistribution does not add them.

### Release history, verbatim from the ARMBench website

> **June, 2026:** New benchmark dataset released for Blade Insertion Affordance Prediction, with ~29,000
> cycles of RGBD images and 3D motion trajectories.
>
> **August, 2025:** New benchmark dataset released for Stow Success Prediction, adding a fourth key
> challenge to ARMBench.
>
> **March, 2023:** Benchmark datasets publicly released for 3 key perception challenges in robotic
> manipulation: (a) Object segmentation, (b) Object identification, and (c) Defect detection.
>
> **January, 2023:** The ARMBench research paper has been accepted for presentation at ICRA, 2023.

## Considerations for Using the Data

**Intended use.** Research and benchmarking of video-based defect detection for robotic manipulation.

**Scope limits.** The videos come from one class of Amazon Robotics workcell in Amazon fulfillment
centres, under that facility's lighting, camera placement and object mix. Models trained here should not
be assumed to transfer to other workcells, camera geometries or inventory distributions without
re-evaluation. Multi-pick defects are excluded from this subset because they are not observable in the
video view; a model trained here cannot detect them.

**Class imbalance.** Defects are rare by design in the operation being measured (4.01% of all 104,075
videos). Accuracy is a poor metric; report per-class precision/recall or average precision, and report
them on the official test split, whose 11.1% defect rate is the realistic one.

**Personal and sensitive information.** The videos show items being handled inside the robotic workcell.
Amazon's website does not publish a de-identification statement for this subset, and this redistribution
does not add one. If you find content that should not be public, please contact
armbench-support@amazon.com.

## Licence and attribution

Quoted verbatim from the ARMBench website's Credits section:

> These datasets are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0). Use
> the following citations if you use these datasets for publication.

| | |
|---|---|
| Licence | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) (`cc-by-4.0`) |
| Original data credit | Amazon.com, Inc. - ARMBench, <http://armbench.com> |
| Modifications to the media | **None.** `archive/armbench-defects-video-0.1.tar.gz` is byte-for-byte identical to Amazon's release object and verifies against Amazon's published SHA-256. |
| Additions | The `labels/` Parquet tables and the derived files in `annotations/` marked "derived" in the file table above were produced from the archive's own annotation files for this redistribution. One file was renamed: `partial_tar_listing.txt` to `annotations/archive_listing_verbose.txt`. |
| Further restrictions | None are applied. This redistribution adds no terms beyond CC BY 4.0. |

Attribution is a binding condition of CC BY 4.0. If you use this data, cite the paper below.

## The ARMBench suite on the Hub

ARMBench is distributed as six downloads, each published as its own dataset under the
[`amazon`](https://huggingface.co/amazon) organization so that you can take only the subset you need.
Every one carries the `armbench` tag, so [`datasets?other=armbench`](https://huggingface.co/datasets?other=armbench)
lists the whole family, and they are grouped in the **ARMBench** collection on the organization page.

| Dataset | Program | Task | Download size |
|---|---|---|---|
| [`amazon/armbench-segmentation`](https://huggingface.co/datasets/amazon/armbench-segmentation) | Sparrow | Object segmentation in clutter | 18.5 GiB |
| [`amazon/armbench-object-identification`](https://huggingface.co/datasets/amazon/armbench-object-identification) | Sparrow | Open-set object identification | 101.6 GiB |
| [`amazon/armbench-defect-image`](https://huggingface.co/datasets/amazon/armbench-defect-image) | Sparrow | Defect detection, images | 65.7 GiB |
| [`amazon/armbench-defect-video`](https://huggingface.co/datasets/amazon/armbench-defect-video) **(this dataset)** | Sparrow | Defect detection, videos | 254.0 GiB |
| [`amazon/armbench-stow-success-prediction`](https://huggingface.co/datasets/amazon/armbench-stow-success-prediction) | Vulcan Stow | Stow success prediction | 37.4 GiB |
| [`amazon/armbench-blade-insertion-affordance`](https://huggingface.co/datasets/amazon/armbench-blade-insertion-affordance) | Vulcan Stow | Blade insertion affordance | 555.7 GiB |

The Sparrow subsets are cited with `mitash2023armbench` (below); the two Vulcan Stow subsets are cited
with `hudson2025stow` — see each dataset's own card for its correct citation.

## Citation

```bibtex
@article{mitash2023armbench,
  title     = {ARMBench: An object-centric benchmark dataset for robotic manipulation},
  author    = {Mitash, Chaitanya and Wang, Fan and Lu, Shiyang and Terhuja, Vikedo
               and Garaas, Tyler and Polido, Felipe and Nambi, Manikantan},
  journal   = {arXiv preprint arXiv:2303.16382},
  year      = {2023}
}
```

This is Amazon's published BibTeX entry with one character removed: the version on armbench.com has a
stray comma after `Vikedo`, which makes BibTeX parse the name under its three-part `last, jr, first`
rule and drop Vikedo Terhuja's given name from the rendered reference. The seven authors and their order
are unchanged and match the arXiv metadata for 2303.16382. No `@inproceedings` ICRA entry is given
because arXiv carries no journal reference for this paper and the proceedings citation could not be
verified.

## Contact

Quoted verbatim from the ARMBench website:

> For questions or to provide feedback, please reach out to us at armbench-support@amazon.com

(c) Amazon.com