Datasets:

amazon
/

armbench-defect-image

---
license: cc-by-4.0
language:
- en
pretty_name: ARMBench Defect Detection - Image Subset
tags:
- armbench
- amazon-robotics
- robotic-manipulation
- defect-detection
- image-classification
- warehouse-automation
task_categories:
- image-classification
size_categories:
- 100K<n<1M
annotations_creators:
- expert-generated
source_datasets:
- original
---
# ARMBench - Defect Detection, Image Subset

This repository is the official Amazon redistribution, on the Hugging Face Hub, of the **image 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 split and metadata files alongside it are Amazon's own
release files, copied out of the bucket unmodified so that the official train/test lists can be read
without downloading 65.7 GiB.

It is the sibling of [`amazon/armbench-defect-video`](https://huggingface.co/datasets/amazon/armbench-defect-video).
The two together are ARMBench's defect detection task; the cards are written as a matched pair.

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.
> Every quotation in this card was taken on 2026-10-28 from the live pages served by armbench.com's own
> CloudFront origin (`/index.html` and `/defects.html`), not from a cached or archived copy.

## 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

> **Image defect detection (66G):** This dataset comprises 13,303 images of objects with defects taking
> through multiple view-points (Transfer-images). For image defection detection, multi-pick and
> package-defects are the two defect classes. 100,000 images of objects with no defects are available in
> the dataset.
>
> **Train-Test-Split:** List of training and test cases.

And from the ARMBench home page, the one-line summary of the defect detection task as a whole, covering
both this subset and the video subset:

> **Defect Detection**
>
> 19,000+ images and 4,000+ videos for rare, but costly, robot-induced defects such as
> multi-pick and packaging defects. Dataset contains 100,000+ no defect cases.

Three editorial notes on the quotations above, so that the differences from the live page are auditable
and not mistaken for transcription errors. First, in the task paragraph, parentheses and italics were
added to `...separated into multiple parts (*deconstruction*)`; the live page italicises the term but
omits the parentheses. Second, **"image defection detection" is the site's own spelling** and is
reproduced here unchanged; it is not a typo introduced by this redistribution. Third, the live page's
sentence about video ("Video defect detection (255G): ...") belongs to the video subset and is not quoted
here. Nothing else was changed, added or reordered.

### What distinguishes this subset from the video subset

**Both defect classes are present here.** The task defines two defect types, multi-pick and
package-defect. In the video subset Amazon states that "Multi-picks are not observable in videos and are
excluded from this dataset", so a model trained there cannot detect them. In this image subset the site
states that "multi-pick and package-defects are the two defect classes". If you need multi-pick, this is
the subset that has it.

The imagery is the *Transfer-images* view: still frames of the object captured from multiple viewpoints
while it is being transferred, rather than a single clip of the transfer.

## Dataset Structure

### What this repository contains

| Path | Bytes | Description |
|---|---|---|
| `archive/armbench-defects-image-0.1.tar.gz` | 70,571,002,037 | Amazon's original release archive, unmodified |

Every byte count above is the exact size of the object as released by Amazon. The sizes of
Amazon's `train.csv`, `test.csv` and `defect_detection_images.ckpt` release objects were measured
on two independently retrieved copies of each, which are MD5-identical to one another.

**There is no derived Parquet label table in this repository, and therefore no dataset viewer.** The
video subset ships one because its labels are small, self-describing per-clip JSON files. The per-image
labels for this subset were not extracted for this redistribution, so rather than ship a `configs:`
block pointing at a file that does not exist - which breaks the viewer and is worse than having none -
this card ships no viewer configuration. The authoritative labels are inside the archive. If a label
table is added later it will appear in the file table above.

### Archive contents

The archive is Amazon's `armbench-defects-image-0.1.tar.gz`, a single gzipped tar. Its root holds the
Transfer-images for each pick together with their annotations and the official split lists; the same
`train.csv` and `test.csv` that Amazon publishes alongside the archive are the split
lists for this subset.

Two practical warnings carried over from the sibling video subset, both of which apply to any
`tar.gz` of this shape:

- **A gzip stream is not seekable.** Any full listing or extraction is a single sequential pass over all
  65.7 GiB. There is no way to pull one image out cheaply.
- **Do not infer the archive's contents from `tar tzf ... | head`.** In the video archive the annotation
  files sit at the end of the gzip stream, so a truncated listing makes the archive look unlabelled when
  it is not. Read a listing to completion before concluding something is missing.

If you only need the official split membership, read Amazon's `train.csv` and Amazon's `test.csv`
and never touch the archive.

### Splits

Amazon's split files are bare, headerless, single-column lists of pick identifiers - one 21-digit ID per
line, no label column. **The lines are CRLF-terminated (`\r\n`), including the last line, and the IDs are
zero-padded**, so read them as strings and strip the carriage return: a naive
`open(path).read().split("\n")` leaves a trailing `\r` on every ID and every join against it then fails
silently, and parsing an ID as an integer destroys its leading zeros. `csv.reader`, or
`pandas.read_csv(path, header=None, dtype=str)`, handles both. Measured from the files in this repository:

| Split | Pick IDs | File | Bytes |
|---|---|---|---|
| train | 6,858 | Amazon's `train.csv` | 157,734 |
| test | 3,990 | Amazon's `test.csv` | 91,770 |
| **total** | **10,848** | | **249,504** |

Every ID is distinct within its file and **the intersection of train and test is empty** (0 IDs in
common), so the split is clean and can be used as published.

The split is at the **pick** level, not the image level: one pick contributes several Transfer-images
from different viewpoints, so the number of images is a multiple of the number of picks. Splitting at the
pick level is what prevents two viewpoints of the same physical object from landing on opposite sides of
the split. **If you build your own split, split on `pick_id`, never on the image filename.**
Amazon's split files carry no class labels, and this redistribution did not enumerate the per-image
classes out of the 65.7 GiB archive. The per-class counts for the official split are therefore not
reported here rather than guessed; the site's own published totals are in the table below.
## Published figures versus measured figures
Amazon's published figures are given first; figures measured from the files in this repository follow,
labelled as such. Both are reported rather than silently reconciled.
| Quantity | Amazon's published figure | Measured here |
|---|---|---|
| Defect images | **13,303** (task page) vs **19,000+** (home page) - *the site's two pages disagree with each other* | not enumerated (see note below) |
| No-defect images | 100,000 (task page), 100,000+ (home page) | not enumerated |
| Defect classes | multi-pick and package-defect (task page) | not enumerated from the archive; both names appear in the third-party mirror's vocabulary (below) |
| Download size | 66G (task page) | **70,571,002,037 bytes** = 65.724 GiB = 70.571 GB |
| Official split | "List of training and test cases" (no count published) | **10,848 pick IDs** = 6,858 train + 3,990 test, 0 overlap |
Two figures on Amazon's own site disagree: the defect detection task page says **13,303** images of
objects with defects, while the home page's task summary says **19,000+** images. This redistribution
does not reconcile them and does not know which population each counts. Both are reported above exactly
as published. For context only, and explicitly *not* offered as a reconciliation, the third-party mirror
described in the next section counts 18,955 defect stills - but it does so over its own 141,132-still
selection, which is a different and larger population than the 113,303 images the task page describes.
The 66G / 65.724 GiB pair is not a discrepancy: Amazon's "66G" is the archive size rounded in
gibibytes, and the exact object is 65.724 GiB.
## A third-party mirror exists; this is the authoritative release
A community mirror of this subset, `correll/armbench_defect_dataset`, is also on the Hub as 128 Parquet
shards. It is useful - it is streamable and it is labelled - but it is **not** the official release and
it should not be used for benchmark numbers, for one decisive reason:
**it does not preserve Amazon's train/test split.** Measured by joining the mirror's `pick_id` column
against Amazon's own split lists:
| Measurement | Value |
|---|---|
| Mirror rows (stills) | 141,132 |
| Mirror distinct `pick_id` | 36,898 |
| Of Amazon's 10,848 official-split picks, present in the mirror | 10,709 |
| Of Amazon's 10,848 official-split picks, **absent** from the mirror | **139** |
| Mirror picks that are **outside** Amazon's official split | **26,189** |
So the mirror is simultaneously missing 139 official picks and carrying 26,189 picks that the official
benchmark does not evaluate on. A result computed on the mirror's own split is not comparable to a result
computed on Amazon's. If you want comparable numbers, take the split from Amazon's own `train.csv` /
`test.csv` release objects.
The mirror's label vocabulary, for orientation (these counts are over the **mirror's** 141,132 stills
and are mirror-derived, not Amazon-published):
| Mirror `label` | Stills |
|---|---|
| `nominal` | 122,177 |
| `package_defect` | 11,142 |
| `multi_pick` | 7,813 |
with `sublabel` taking nine values - `none` (130,145), `open_book_jacket` (3,056), `open_box` (2,880),
`book_jacket` (2,600), `open_book` (1,372), `partial_box` (665), `torn_bag` (249), `crush_box` (89),
`empty_bag` (76). Note that these sublabel names are finer-grained than the two subclasses Amazon's task
page defines (*open* and *deconstruction*), and that 155 `package_defect` stills carry `sublabel = none`.
Do not assume the mirror's vocabulary is Amazon's.
## The benchmark target
The task, in Amazon's words, "is to identify if a robotic manipulation activity resulted in a defect",
with "multi-pick and package-defects" as "the two defect classes" for this subset. Because a single
activity can in principle exhibit more than one defect type, and because the sibling video subset's
released baseline uses two independent sigmoid channels rather than a softmax over classes, treat the
target as **per-class presence**, not as mutually exclusive classes, unless you have verified otherwise
on the labels inside the archive.
Amazon releases a baseline for this subset as `defect_detection_images.ckpt` (282,633,127
bytes). This card does not report its accuracy: no benchmark number for the image baseline is published
on the website, and none was measured here.
## 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
# The repository holds exactly one data file: Amazon's release archive.
hf download amazon/armbench-defect-image \
  --repo-type dataset \
  --include 'archive/armbench-defects-image-0.1.tar.gz' \
  --local-dir ./armbench-defect-image
# Amazon publishes this SHA-256 for the archive (quoted from its sidecar in the
# release bucket; the sidecar itself is not mirrored here, so write the line out):
cd ./armbench-defect-image/archive
echo '5072905775299a6a2ba2fded9eda9009149c2e0d848493d34ff6b0086f8699d8  ./armbench-defects-image-0.1.tar.gz' \
  | sha256sum -c -
```
The sidecar's single line is, verbatim:
```
5072905775299a6a2ba2fded9eda9009149c2e0d848493d34ff6b0086f8699d8 ./armbench-defects-image-0.1.tar.gz
```
Note the `./` path prefix: run `sha256sum -c` from the directory that contains the archive, or the check
will report the file as missing.
Splits only, no archive - a few hundred kilobytes instead of 65.7 GiB:
```text
The official split lists are not mirrored as separate files in this repository.
Amazon ships them inside the release archive (the site's own "Train-Test-Split:
List of training and test cases"), with 6,858 train and 3,990 test pick ids.
```
Or to try the data before committing to the full download, take the shared 66 MiB sample
`armbench-defects-tiny.tar.gz`.
## 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
Defect annotations were produced by human annotators. Amazon's website does not publish the annotation
protocol, the annotator population, the adjudication procedure, or inter-annotator agreement for this
subset, and this redistribution does not add them. Where a number is not published, this card says so
rather than estimating it.
### 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.
This subset is part of the March 2023 release.
## Considerations for Using the Data
**Intended use.** Research and benchmarking of image-based defect detection for robotic manipulation.
**Scope limits.** The images 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.
**Class imbalance.** Defects are rare by design in the operation being measured: against 100,000
no-defect images the site reports 13,303 defect images, so the defect rate is of order ten percent and
the published no-defect pool alone is large enough to make accuracy meaningless as a metric. Report
per-class precision/recall or average precision, and report them on the official test split.
**Pick-level leakage.** Several Transfer-images come from the same pick. Any custom split, augmentation
scheme or deduplication must be done at the `pick_id` level; splitting at the image level leaks the same
physical object across train and test and will inflate your score.
**Personal and sensitive information.** The images 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.

## Files in this repository (generated from the uploaded tree)

| Path | Bytes |
|---|---|
| `License-CC-BY.txt` | 18,657 |
| `annotations/test.csv` | 91,770 |
| `annotations/train.csv` | 157,734 |
| `archive/armbench-defects-image-0.1.tar.gz.sha256` | 100 |
| `archive/armbench-defects-tiny.tar.gz` | 69,609,824 |
| `baseline/defect_detection_images.ckpt` | 282,633,127 |

The large release archive is listed in the sections above; the files here are the small
companion objects (Amazon's published checksum sidecars, the tiny sample archives, the
official split lists and the released baseline checkpoints), all unmodified.

## 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-image-0.1.tar.gz` is byte-for-byte identical to Amazon's release object and verifies against Amazon's published SHA-256. |
| Additions | Only this card. The one data file, `archive/armbench-defects-image-0.1.tar.gz`, is an unmodified Amazon release object under an `archive/` prefix; **no file was renamed** and no derived table, label file or checkpoint is added. Amazon's sidecar, sample archive, split lists and baseline checkpoint are described in this card but are not mirrored here. |
| 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) **(this dataset)** | Sparrow | Defect detection, images | 65.7 GiB |
| [`amazon/armbench-defect-video`](https://huggingface.co/datasets/amazon/armbench-defect-video) | 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