amazon/armbench-segmentation

ARMBench - Object Segmentation

This repository is the official Amazon redistribution, on the Hugging Face Hub, of the object segmentation subset of ARMBench. The media payload is Amazon's original release archive, byte-for-byte. Nothing in it is reprocessed, and no derived label tables are added - the annotations ship inside the archive itself.

This one download carries three official subsets - Mix-Object-Tote, Zoomed-Out-Tote-Transfer-Set and Same-Object-Transfer-Set - which exist precisely so that cross-subset transfer can be measured.

Project home: http://armbench.com  |  Paper: arXiv:2303.16382  |  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. Direct URLs to individual task pages under http://armbench.com return 404; reach the task pages from the site's own navigation instead.

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 that download; the others are published under the same amazon organization with the armbench- prefix.

The site's one-line summary of this task, from the home page, verbatim. On the page this sentence is the body of the home-page card headed Object Segmentation; that heading is a separate element and carries no colon, so it is kept outside the quotation:

450,000+ high-quality labels for object segments on 50,000+ images. Clutter and variety of objects present a novel challenge for instance segmentation algorithms.

The object segmentation task, verbatim

Quoted verbatim from the site's Object Segmentation task page (segmentation.html), retrieved 2026-10-28:

Instance segmentation is a computer vision technique utilized to detect and separate individual objects in an image. Within the realm of warehouse automation, instance segmentation is used to identify and define distinct objects that are stored in containers. The outcomes of instance segmentation can be used to provide information to subsequent robotic processes, such as the identification of objects and generation of grasping strategies. It is an essential factor in achieving high-efficiency robot manipulation as it facilitates more precise and accurate object detection and handling.

The three subsets, verbatim

Still quoting the same task page:

We have divided the Object Segmentation dataset into three distinct subsets to showcase the challenge of transferring learned knowledge across tasks. The subsets are as follows:

Mix-Object-Tote (14G): This subset consists of close-up images of mixed objects that are stored in either yellow or blue totes. Mix-Object-Tote comprises a total of 44,253 images of size 2448 by 2048 pixels and 467,225 annotations, with an average of 10.5 instances per tote.

Zoomed-Out-Tote-Transfer-Set (1.5G): This subset includes mixed objects placed in a yellow tote that were captured with sensors positioned further away from the tote, under different lighting conditions. The Zoomed-Out-Tote-Transfer-Set contains 5,837 images of size 2046 by 2046 pixels and 43,401 annotations, with an average of 7.5 objects per tote.

Same-Object-Transfer-Set (3G): This subset consists of multiple same objects placed in close proximity within various storage units. The Same-Object-Transfer-Set comprises 3,323 images of size 2048 by 1500 pixels and 12,664 annotations, with an average of 3.8 objects per scene.

These labels are provided per-image in the format that is compatible with LabelMe, as well as a train, validation, and test split format that is compatible with MS-COCO format.

One editorial note, so that the differences from the live page are auditable and not mistaken for transcription errors: the subset names and the sizes in parentheses are rendered in bold here; on the live page they are run-in headings of each paragraph. No word was changed, added or reordered.

Why three subsets. The site states the purpose outright - "to showcase the challenge of transferring learned knowledge across tasks". Mix-Object-Tote is the large in-domain set. The other two are transfer sets: Zoomed-Out-Tote-Transfer-Set changes the camera standoff and the lighting while keeping the mixed-object premise, and Same-Object-Transfer-Set changes the content premise entirely to multiple copies of the same object in close proximity, which is the hard case for instance separation. Training on mix-object-tote and reporting on the two transfer sets is the intended use. Pooling all three into one training set destroys the measurement the release was built to support.

Subset figures, as published by Amazon

Subset	Images	Image size (px)	Annotations	Mean instances per image	Site download size
Mix-Object-Tote	44,253	2448 x 2048	467,225	10.5 per tote	14G
Zoomed-Out-Tote-Transfer-Set	5,837	2046 x 2046	43,401	7.5 per tote	1.5G
Same-Object-Transfer-Set	3,323	2048 x 1500	12,664	3.8 per scene	3G
Total	53,413	-	523,290	-	18.5G

The totals row is arithmetic on Amazon's own three figures; the site does not print a total of its own. Both the images total and the annotations total are consistent with the home page's rounded lower bounds of "50,000+ images" and "450,000+ high-quality labels".

Dataset Structure

What this repository contains

Path	Bytes	Description
archive/armbench-segmentation-0.1.tar.gz	19,814,367,023	Amazon's original release archive, unmodified
README.md	-	This card

That is the whole repository: one media file and this card. The byte count is exact, not rounded - it is the ContentLength of Amazon's own release object, and the uploaded file is byte-identical to it.

Other objects in Amazon's release that are not redistributed here

Amazon's segmentation release contains three further objects. They are listed for completeness, with their exact sizes, so that nobody mistakes their absence for an incomplete mirror - but none of them is in this repository, and a hf download for any of these paths will fail:

Amazon release object	Bytes	Why it is not here
armbench-segmentation-tiny.tar.gz	1,564,893	A five-image sample (described below). Not mirrored; the full archive supersedes it.
armbench-segmentation-0.1.zip	19,854,049,522	The same contents as the .tar.gz. Mirroring both would duplicate ~18.5 GiB of identical image bytes.
latest.pt	351,086,890	A baseline checkpoint. Amazon's website documents nothing about its architecture, training recipe or metrics, so it is not redistributed with claims this card cannot support.

Amazon publishes no SHA-256 sidecar for the segmentation release, so none is mirrored either.

Nothing is missing from the .tar.gz: it carries all three subsets and their annotations.

Archive layout

Everything sits under a single top-level directory, armbench-segmentation-0.1/, with one directory per official subset. Verified directly from the release archive's own member table:

armbench-segmentation-0.1/
  mix-object-tote/
    images/<IMAGE_ID>.jpg          # 10-character uppercase alphanumeric IDs
    ...

Image members carry 2022-08-20 modification times. The archive's root directory entry carries 2023-03-23, while its mix-object-tote/ and images/ entries carry 2023-03-07 - consistent with the March 2023 public release.

Partial member census, measured

A full member-by-member listing of the archive was deliberately not completed for this card. The machine that staged this release was shared, and a complete tar -tzf pass over 19,814,367,023 bytes of already-compressed JPEG is one long sequential read that would have competed with the staging transfer. The census was stopped after the first 28,131 members. What it does establish:

Measured property	Value	Status
Archive root directory	armbench-segmentation-0.1/	confirmed
Subset directory observed	mix-object-tote/	confirmed
Image path layout	<subset>/images/<IMAGE_ID>.jpg, IDs 10 uppercase alphanumerics	confirmed for mix-object-tote
Image members enumerated under mix-object-tote/images/	28,131 of the 44,253 Amazon publishes	lower bound; enumeration incomplete
Bytes in those members	9,845,784,837	lower bound
Mean image size	~350 kB	over the members enumerated

Two things this card therefore does not claim: the on-disk directory names of the Zoomed-Out-Tote-Transfer-Set and Same-Object-Transfer-Set subsets, and the exact filenames of the LabelMe and MS-COCO annotation files. Those members sit later in the gzip stream than the census reached. The annotation formats are the ones Amazon's task page states and are quoted above; the annotation paths are not reproduced here because they were not read.

The sample archive, measured

armbench-segmentation-tiny.tar.gz (1,564,893 bytes) is part of Amazon's release but is not redistributed in this repository; it was listed in full for this card. It contains a single directory armbench-segmentation-tiny/ holding five JPEG images and nothing else: P2S0835EOA.jpg, 9TARTKRF1X.jpg, EZ5895NGB8.jpg, LO95CK8OLH.jpg, OAR1KXNIME.jpg.

The sample archive carries no annotations. It is a 1.5 MB look at the imagery, useful for checking resolution and visual domain before committing to the full download. It is not a runnable mini-benchmark, and it is not laid out like the full archive - its images sit at the root of its own directory rather than under a <subset>/images/ path. If you need labels, you need the full archive.

Annotation formats

The site specifies two annotation formats, and they serve different purposes:

Format	Granularity	What it is for
LabelMe	One file per image	Per-image polygon instance masks; the annotation of record
MS-COCO	One file per split (train / validation / test)	The official train / validation / test partition, in the format most detectors and the pycocotools evaluator consume directly

Note that this task has three splits - train, validation and test - unlike the ARMBench defect detection subsets, which publish only train and test. Report on test; tune on validation.

No dataset-viewer configuration

This card intentionally declares no configs: block, so the Hub shows no dataset viewer for this repository. The release ships JPEG images with LabelMe and MS-COCO JSON annotations, not Parquet or CSV; a configs: block pointing at a file that does not exist would break the viewer rather than populate it. Load the archive directly, or read the MS-COCO split JSONs with pycocotools.

Published figures versus measured figures

Amazon's published figures are given first; 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
Images	50,000+ (home page); 44,253 + 5,837 + 3,323 per subset (task page)	53,413 by summing the task page's three per-subset figures
Object segment labels	450,000+ (home page); 467,225 + 43,401 + 12,664 per subset (task page)	523,290 by summing the task page's three per-subset figures
Download size	14G + 1.5G + 3G = 18.5G (task page)	19,814,367,023 bytes = 18.454 GiB = 19.81 GB, measured on the staged release object
Sample archive size	not published	1,564,893 bytes, measured

The home page's "450,000+" and "50,000+" are rounded-down lower bounds and do not contradict the per-subset figures. The one difference worth noting is the download size: the task page's per-subset sizes are decimal-gigabyte approximations that sum to 18.5G, while the single distributed archive is 18.454 GiB. These describe the same bytes at different precision.

Verifying what you downloaded

The archive in this repository is byte-identical to Amazon's release object, so a byte-count check is a meaningful first test:

hf download amazon/armbench-segmentation \
  archive/armbench-segmentation-0.1.tar.gz \
  --repo-type dataset --local-dir ./armbench-segmentation

stat -c %s ./armbench-segmentation/archive/armbench-segmentation-0.1.tar.gz
# expected: 19814367023

This card does not quote a SHA-256 for the segmentation archive, and no .sha256 file is present in this repository. Amazon publishes checksum sidecars only for the two defect-detection subsets; there is no published checksum for the segmentation release, so there is none to mirror and inventing one would be worse than omitting it. The byte count above is the integrity check this redistribution can honestly offer.

Peek at the contents without extracting the whole thing - the gzip stream is not seekable, so a full listing or extraction is a single sequential pass:

tar -tzf armbench-segmentation-0.1.tar.gz | head -20

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

Instance masks were produced by human annotators. Amazon's website does not publish the annotation protocol, the annotator population, the mask quality-control procedure, or inter-annotator agreement for this subset, and this redistribution does not add them. The site's only quality claim is the home page's phrase "high-quality labels". The site likewise does not document the architecture, training recipe or reported metrics of the baseline checkpoint shipped with this release.

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.

Object segmentation is one of the three March 2023 releases.

Considerations for Using the Data

Intended use. Research and benchmarking of instance segmentation in cluttered containers, and of cross-domain transfer between the three subsets.

Do not pool the subsets. The three subsets were separated, in Amazon's own words, "to showcase the challenge of transferring learned knowledge across tasks". Merging them into one training pool and reporting a single score discards the only property this release was designed to expose. If you report benchmark numbers, report them per subset.

Scope limits. The images come from Amazon Robotics workcells in Amazon fulfillment centres, under those facilities' lighting, camera placement and inventory mix, and the containers are yellow or blue totes and warehouse storage units. Models trained here should not be assumed to transfer to other container types, camera geometries or object distributions without re-evaluation - which is, in miniature, exactly what the two transfer subsets measure.

Resolution and cost. These are large images - up to 2448 x 2048 - and the mean scene holds 10.5 instances. Expect decoding, not GPU time, to dominate a naive data loader. Also note that the three subsets have three different native resolutions and aspect ratios; a fixed resize that suits Mix-Object-Tote will distort Same-Object-Transfer-Set.

Class labels. This is instance segmentation of objects in clutter. The site's description is in terms of object segments and instances per tote, not of a semantic class vocabulary; do not assume a class taxonomy is present without checking the annotation files.

Personal and sensitive information. The images show inventory items inside robotic workcells. 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
archive/armbench-segmentation-tiny.tar.gz	1,564,893
baseline/latest.pt	351,086,890

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 (cc-by-4.0)
Original data credit	Amazon.com, Inc. - ARMBench, http://armbench.com
Modifications to the media	None. archive/armbench-segmentation-0.1.tar.gz is byte-for-byte identical to Amazon's release object, at 19,814,367,023 bytes.
Additions	Only this card. No derived label tables are published for this subset; the annotations are read from the archive itself.
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 organization so that you can take only the subset you need. Every one carries the armbench tag, so 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 (this dataset)	Sparrow	Object segmentation in clutter	18.5 GiB
amazon/armbench-object-identification	Sparrow	Open-set object identification	101.6 GiB
amazon/armbench-defect-image	Sparrow	Defect detection, images	65.7 GiB
amazon/armbench-defect-video	Sparrow	Defect detection, videos	254.0 GiB
amazon/armbench-stow-success-prediction	Vulcan Stow	Stow success prediction	37.4 GiB
amazon/armbench-blade-insertion-affordance	Vulcan Stow	Blade insertion affordance	555.7 GiB

The Sparrow subsets, including this one, 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

@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 and the field indentation reflowed (the site prints title = {...}; it is aligned here). 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. All 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