amazon/armbench-blade-insertion-affordance

ARMBench - Blade Insertion Affordance Prediction

This repository is the official Amazon redistribution, on the Hugging Face Hub, of the Blade Insertion Affordance Prediction subset of ARMBench. The media payload is Amazon's original release archive, byte-for-byte; because that single file is larger than the Hub's hard per-file ceiling it is stored as six byte-exact parts that cat back into the original object (see Reassembling the archive). No pixel, pose or trajectory value has been altered.

Project home: http://armbench.com  |  Paper: arXiv:2505.04572  |  Contact: armbench-support@amazon.com

Note on the project website. http://armbench.com is Amazon's canonical home for ARMBench. 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.

Note on the citation. This subset comes from the Vulcan Stow program, not from the Sparrow program described in the original ARMBench ICRA paper. The correct citation is therefore hudson2025stow (arXiv:2505.04572), reproduced in full with all 37 authors in Citation. Attribution is a binding condition of CC BY 4.0; please do not abbreviate the author list.

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 Vulcan Stow group, described on the site as:

Vulcan Stow - Amazon's first robot with a sense of touch, built on advances in robotics, engineering, and physical AI. Vulcan can manipulate objects within fabric-covered pods that are divided into compartments, each holding up to 10 items. Using specialized end-of-arm tooling and force feedback sensors, Vulcan can detect when it makes contact with objects and apply appropriate force, making stowing operations more efficient and safer while achieving human levels of packing density and speed.

The site's one-line summary of this task, from the home page, verbatim:

Blade Insertion Affordance - ~29,000 blade insertion cycles with RGBD images and 3D motion trajectories. Predict blade insertion affordances across three behavior types.

ARMBench documents five benchmark tasks and distributes them as six downloads. This repository is one of those six; the others are published under the same amazon organization with the armbench- prefix.

The blade insertion affordance task, verbatim

Quoted verbatim from the task page on armbench.com:

Blade Insertion Affordance Prediction

The Blade Insertion Affordance Prediction task focuses on predicting where and how to insert a blade into a bin given visual observations. In Amazon Vulcan Stow workcells, a blade is deployed into fabric-covered pods to create space for item placement. This dataset provides RGBD images, camera calibration, and full 3D motion trajectories for three distinct insertion strategies.

The Blade Insertion Affordance dataset contains ~29,000 blade insertion cycles with data collected from Amazon Vulcan Stow workcells. The dataset supports predicting 3D blade insertion affordances: sparse key SE(3) poses and blade extension lengths, given bin observations.

Task: Given a bin observation O ∈ {RGB image, depth image}, predict the 3D blade insertion affordance, a set of K sparse key SE(3) poses P_k and blade extension lengths b_k that define where and how to insert the blade.

Editorial note on the quotation above: the live page renders the pose and length symbols as Pk and bk with typographic subscripts; they are reproduced here as P_k and b_k. The page's O ∈ {RGB image, depth image} is reproduced exactly. Nothing else was changed, added or reordered. The page also carries a link captioned "See How Amazon's Vulcan robots use touch to plan and execute motions for an overview of blade insertion behaviors" and nine example bin images (pre-stow, mid-stow and post-stow for each strategy), which are not reproduced here.

Insertion strategies, verbatim

The page lists three strategies, each with its cycle count and its sequence of action primitives:

Insertion Strategies:

• Direct Insert (14,948 cycles): APPROACH → EXTEND_BLADE
• Corner Wall Insert (8,671 cycles): APPROACH → EXTEND_BLADE → WALL_SEEK → LIP_SEEK
• Engage Wall Insert (5,887 cycles): APPROACH → ITEM_PUSH → PREPARE_TO_SEEK_WALL → WALL_SEEK → FOLLOW_WALL → EXTEND_BLADE

Those three counts sum to 29,506 cycles, which is the exact figure behind the site's "~29,000". The strategy mix is therefore:

Strategy	insert_strategy value	Cycles	Share	Action primitives
Direct Insert	DIRECT_INSERT	14,948	50.66%	2
Corner Wall Insert	CORNER_WALL_INSERT	8,671	29.39%	4
Engage Wall Insert	ENGAGE_WALL_INSERT	5,887	19.95%	6
total		29,506	100%

The strategies are not interchangeable and the class mix is not balanced. The number of action primitives, and therefore the number of key poses K a model must predict, differs per strategy: 2, 4 and 6. A model that assumes a fixed K will be wrong on two of the three strategies. The site does not publish a train/test split for this subset (see What the site does not publish), so if you report numbers, state your own split construction and stratify it by insert_strategy.

Data per cycle, verbatim

Data per cycle:

• Bin Images (RGB): Cropped bin images at pre-stow, mid-stow, and post-stow stages (JPEG)
• Bin Depth Images: Corresponding 16-bit PNG depth images (values in mm)
• Full-resolution Stereo: Primary and secondary camera images (2448×2048)
• 3D Motion Trajectories: Blade tip SE(3) poses and arm trajectories per action primitive
• Camera Calibration: Intrinsic matrix, distortion coefficients, and camera-to-bin SE(3) transform

Field layout, verbatim

The table below is Amazon's own feature table from the task page, transcribed verbatim:

Category	Feature	Description	Format
Images	bin_images/*.jpg	Cropped bin RGB at pre/mid/post stow stages	JPEG
	bin_depth_images/*.png	Depth images (divide by 1000 for meters)	16-bit PNG
	images_raw/*/primary/*.jpg	Full-resolution stereo camera images	JPEG (2448×2048)
Camera	cam_H_bin	SE(3) transform from camera to bin frame	4×4 matrix
	camera_matrix	Intrinsic matrix [fx, 0, cx; 0, fy, cy; 0, 0, 1]	3×3 matrix
	distortion	Lens distortion coefficients	float array
Trajectories	blade_translation	Blade tip xyz position over time	(N, 3) array
	blade_rot	Blade tip quaternion [x, y, z, w] over time	(N, 4) array
	arm_translation / arm_rot	Arm end-effector pose over time	list of xyz + quat
Labels	insert_strategy	DIRECT_INSERT / CORNER_WALL_INSERT / ENGAGE_WALL_INSERT	string
	action_sequence	Segmented action primitives with per-action trajectories	list of dicts

Two things in that table are easy to get wrong and are worth repeating. Depth is in millimetres - divide the 16-bit PNG values by 1000 for metres. blade_rot is [x, y, z, w], scalar-last; SciPy's Rotation.from_quat expects scalar-last and will accept it directly, while many other libraries (including pytorch3d and ROS tf.transformations' quaternion_matrix convention in some wrappers) expect scalar-first and will silently produce a wrong rotation.

Amazon's table describes images_raw/*/primary/*.jpg only. The task page's prose says "Primary and secondary camera images", so a secondary sibling of primary is implied, but the site's table does not list its path and this card does not invent one.

Dataset Structure

What this repository contains

Path	Bytes	Description
archive/blade_insertion.zip.part000	100,000,000,000	Part 1 of 6 of Amazon's original release archive
archive/blade_insertion.zip.part001	100,000,000,000	Part 2 of 6
archive/blade_insertion.zip.part002	100,000,000,000	Part 3 of 6
archive/blade_insertion.zip.part003	100,000,000,000	Part 4 of 6
archive/blade_insertion.zip.part004	100,000,000,000	Part 5 of 6
archive/blade_insertion.zip.part005	96,637,520,379	Part 6 of 6 (short final part)

The six parts total 596,637,520,379 bytes - the exact size of Amazon's blade_insertion.zip release object. That is 555.662 GiB (596.638 GB), making this the largest of the six ARMBench downloads, over twice the size of the video defect subset.

Why it is split, and what that means for you

Hugging Face enforces a hard per-file ceiling of 500 GB. At 596,637,520,379 bytes blade_insertion.zip exceeds it and cannot be stored as a single file on the Hub under any repository setting. It is therefore stored as six parts produced by GNU split with -d -a 3 (numeric, 3-digit suffixes) and reassembled by concatenation. Splitting is lossless and byte-exact: split writes a plain byte-range partition with no container, header or re-encoding, so concatenating the parts in numeric order reproduces the original object bit-for-bit. The authoritative cross-check against Amazon's object is its byte count, 596,637,520,379, reached by summing the six parts; see Integrity for what can and cannot be verified cryptographically.

Editorial note on the split command. GNU coreutils reads split -b 100G as 100 GiB (107,374,182,400 bytes) and split -b 100GB as 100 GB (100,000,000,000 bytes). The part boundaries in the table above are the decimal-GB ones, and they sum exactly to 596,637,520,379. Either suffix yields six parts and either is safe, because reassembly is by cat and is boundary-agnostic. If a part's length differs from the table, trust the parts as uploaded: what must hold is that the six lengths sum to 596,637,520,379 and that they are concatenated in numeric order.

Reassembling the archive

hf download amazon/armbench-blade-insertion-affordance \
  --repo-type dataset --include 'archive/*' --local-dir ./armbench-blade-insertion

cd ./armbench-blade-insertion/archive

# 1. check the parts are all present and total the published size before
#    spending the disk. Expect 6 files and 596637520379 bytes.
ls blade_insertion.zip.part* | wc -l          # expect: 6
cat blade_insertion.zip.part* | wc -c         # expect: 596637520379
#    (or, without re-reading the data: du -b --total blade_insertion.zip.part*)

# 2. concatenate in numeric order -- the glob sorts correctly because the
#    suffixes are zero-padded 3-digit numbers (part000 .. part005)
cat blade_insertion.zip.part* > blade_insertion.zip

# 3. verify the reassembled object's size against Amazon's own ContentLength.
#    Amazon publishes no SHA-256 for this object, so size is the check it
#    supports -- see the Integrity section.
stat -c %s blade_insertion.zip                # expect: 596637520379

# 4. only now is it safe to reclaim the parts
rm blade_insertion.zip.part*

unzip -l blade_insertion.zip | head

Budget 1.1 TiB of free disk, not 556 GiB: step 2 writes a second full copy before step 4 can delete the first. If you are short on space, cat the parts straight into unzip is not an option - zip's central directory is at the end of the file, so unzip needs the assembled file to be seekable. If disk is the binding constraint, stream the parts to a filesystem that supports reflink copies, or assemble onto a volume you can afford to fill.

Do not use cat blade_insertion.zip.part0* with a shell whose glob ordering is locale-dependent, and do not reorder the parts; the suffixes are zero-padded precisely so that lexical and numeric order coincide.

Archive contents

The release ships as a single zip. Per the site's field layout, each cycle contributes cropped RGB bin images at three stow stages, matching 16-bit depth PNGs, full-resolution stereo JPEGs at 2448×2048, per-primitive blade-tip and arm trajectories, and camera calibration. Across 29,506 cycles and 596,637,520,379 bytes that averages 20,220,888 bytes (19.28 MiB) per cycle; the full-resolution stereo imagery dominates that figure.

This card does not publish a member-by-member listing of the zip. Producing one requires reading the whole 596 GB object, and no such listing has been generated for this release. If you need one, generate it from your own assembled copy with unzip -l blade_insertion.zip.

A note on the extracted form in Amazon's source bucket

Amazon's source bucket also holds the extracted form of this same data, as loose objects under four month-named directories (05_2025, 06_2025, 07_2025, 08_2025) that sit beside the zip in that bucket; this publishing run's inventory counted 2,444,299 of them. That extracted form is deliberately not republished here: the zip is the distribution artifact, and shipping both would place ~596 GB of identical bytes in the repository twice. Nothing is missing as a result. The site does not publish a statement about what those directory names denote, and this card does not infer one.

No configs: block, and therefore no dataset viewer

This repository intentionally has no configs: block in its front matter and the dataset viewer is correspondingly inactive. No Parquet or CSV table has been derived for this subset, and a configs: block pointing at a path that does not exist breaks the viewer with an error, which is worse for a reader than no viewer at all. load_dataset() will not work here; download and reassemble the archive instead.

Published figures versus measured figures

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

Quantity	Amazon's published figure	Figure from this release
Blade insertion cycles	~29,000 (home page and task page)	29,506, the exact sum of the task page's own three per-strategy counts (14,948 + 8,671 + 5,887)
Strategy naming	"three behavior types" (home page) vs "three distinct insertion strategies" (task page)	The data field is named insert_strategy; this card uses "strategy" throughout
Download size	The task page publishes no size figure for this subset, unlike the defect and segmentation pages	596,637,520,379 bytes = 555.662 GiB = 596.638 GB, read from the release object's own S3 ContentLength; the verifying command and its output are in Integrity
Depth units	"values in mm" / "divide by 1000 for meters"	Consistent; both statements appear on the page and agree

The ~29,000 / 29,506 gap is a rounding on Amazon's part, not a conflict: 29,506 is reconstructed from Amazon's own published per-strategy counts on the same page, so both figures come from the site. It is listed here only so that a reader who counts 29,506 cycles does not think the download is wrong.

Integrity: what Amazon publishes, and what this repository adds

The byte count above is Amazon's own, and you can check it yourself against the release object:

$ aws s3api head-object --bucket armbench-dataset \
    --key vulcan_stow/blade_insertion/blade_insertion.zip
{
    "AcceptRanges": "bytes",
    "LastModified": "2026-06-24T19:09:40+00:00",
    "ContentLength": 596637520379,
    "ETag": "\"6c7277684c0cb6a654ced468c5695101-8891\"",
    "ContentType": "application/zip",
    "ServerSideEncryption": "AES256",
    "Metadata": {}
}

Amazon does not publish a SHA-256 for blade_insertion.zip. Some other ARMBench subsets ship a .sha256 sidecar beside their archive; this one does not, and the release object carries no SHA-256 in its metadata. What Amazon's object does carry is:

Amazon-side value	Literal value	What it is good for
ContentLength	596637520379	The exact size, and the primary cross-check against this repository's six parts
ChecksumCRC64NVME, with ChecksumType: FULL_OBJECT	ZiKBoBdybmo=	A whole-object CRC64, retrievable with s3api get-object-attributes --object-attributes Checksum. Not a cryptographic digest, and not computable with sha256sum
ETag	6c7277684c0cb6a654ced468c5695101-8891	Not a content digest. The -8891 suffix marks a multipart upload, so this is a digest over Amazon's 8,891 upload parts and cannot be reproduced from the assembled file

This repository therefore publishes no SHA-256 either, for the parts or for the reassembled whole: there is no Amazon digest to mirror, and a digest computed here would attest only to this copy, not to Amazon's original. Verification of this subset rests on the byte count - each part's length in the table above, and their sum, 596,637,520,379, matching Amazon's ContentLength - plus the CRC64 above if you hold the original. This card does not claim an Amazon SHA-256 that does not exist, and does not ship a locally-computed one dressed up as an integrity guarantee.

What the site does not publish

Stated explicitly so that these gaps are not mistaken for omissions in this card. For this subset, armbench.com does not publish:

• A train/test split. The other ARMBench subsets ship official split lists; this one does not. Any split you use is your own, and should be stratified by insert_strategy.
• A baseline model or baseline metrics. No checkpoint and no reported accuracy for this task.
• An evaluation metric or scoring protocol. The task is defined ("predict K sparse key SE(3) poses and blade extension lengths") but no pose-error metric, tolerance or leaderboard definition is given.
• A download-size figure for this subset.
• The number of workcells, sites, pods or bins, or how cycles are distributed across them.
• Label provenance in detail. insert_strategy and action_sequence are segmentations of executed robot motion rather than human image annotations, which is why this card's front matter declares annotations_creators: machine-generated; the site does not describe the segmentation procedure, and no human annotation protocol, annotator population or inter-annotator agreement is published.
• A de-identification statement. See Personal and sensitive information below.
• A secondary camera path, despite the prose mentioning secondary camera images.

This redistribution does not add any of the above.

Considerations for Using the Data

Intended use. Research and benchmarking of visual affordance prediction for contact-rich robotic manipulation: predicting where and how to insert a blade into a fabric-covered pod compartment from RGB-D bin observations.

Scope limits. The cycles come from Amazon Vulcan Stow workcells, under those facilities' lighting, camera placement, pod geometry and inventory mix. The blade, its kinematics and the fabric pod are specific hardware. Affordances learned here should not be assumed to transfer to other end effectors, other container types or other camera geometries without re-evaluation.

Only executed cycles are present. Each cycle records a strategy the controller actually chose and carried out. The dataset is therefore a sample from the deployed policy's own distribution, not a uniform or randomised sample over possible insertions, and it contains no counterfactuals: for a given bin there is no record of what the two unchosen strategies would have done. Treat insert_strategy as a logged decision under an unknown policy, not as a ground-truth optimal label, and expect confounding if you train a strategy classifier and interpret it causally.

Class imbalance. The strategy mix is 50.66 / 29.39 / 19.95 percent. Report per-strategy metrics, not a single pooled number.

Coordinate frames matter here more than usual. Poses are expressed with respect to the camera and the bin, and cam_H_bin is the transform between them. Mixing frames, or reading blade_rot as scalar-first, produces plausible-looking numbers that are geometrically wrong. Verify your transform chain on a single cycle by reprojecting the blade tip trajectory into the bin image before trusting it at scale.

Personal and sensitive information. The imagery shows inventory and pod interiors inside a 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.

Dataset Creation

The data was collected from Amazon Vulcan Stow workcells, as described on the task page and in the Vulcan Stow paper, Stow: Robotic Packing of Items into Fabric Pods (arXiv:2505.04572). Beyond the task page text quoted above, the site does not publish a collection protocol for this subset.

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 the June 2026 release, the most recent and the fifth ARMBench benchmark task.

Files in this repository (generated from the uploaded tree)

Path	Bytes
License-CC-BY.txt	18,657

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. The six archive/blade_insertion.zip.part* files concatenate, byte-for-byte, into Amazon's release object: they sum to exactly its 596,637,520,379 bytes, and splitting a file into byte ranges alters no content. Amazon publishes no SHA-256 for this object, so see Integrity for what can and cannot be verified against the original.
Additions	Only this card. The six archive/blade_insertion.zip.part* files are a byte-range partition of Amazon's single release object, produced with GNU split; no checksum file, label table or derived artifact is added, and nothing was removed or renamed.
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 with its complete author list.

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	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 (this dataset)	Vulcan Stow	Blade insertion affordance	555.7 GiB

The four Sparrow subsets are cited with mitash2023armbench; the two Vulcan Stow subsets, this one and stow success prediction, are cited with hudson2025stow below.

Citation

This is the Vulcan Stow citation as published on armbench.com, with all 37 authors reproduced verbatim and in the site's order. Do not shorten this list to "et al." or "and others" in a works-cited entry: attribution is a licence condition, not a formatting preference.

@article{hudson2025stow,
  title     = {Stow: Robotic Packing of Items into Fabric Pods},
  author    = {Hudson, Nicolas and Hooks, Josh and Warrier, Rahul and Salisbury, Curt and
               Hartley, Ross and Kumar, Kislay and Chandrashekhar, Bhavana and Birkmeyer, Paul and
               Tang, Bosch and Frost, Matt and Thakar, Shantanu and Piaskowy, Tony and
               Nilsson, Petter and Petersen, Josh and Doshi, Neel and Slatter, Alan and
               Bhatia, Ankit and Meeker, Cassie and Xue, Yuechuan and Cox, Dylan and
               Kyriazis, Alex and Lou, Bai and Hasan, Nadeem and Rana, Asif and
               Chacko, Nikhil and Xu, Ruinian and Faal, Siamak and Seraj, Esi and Agrawal, Mudit and
               Jamieson, Kevin and Bisagni, Alessio and Samzun, Valerie and Fuller, Christine and
               Keklak, Alex and Frenkel, Alex and Ratliff, Lillian and Parness, Aaron},
  journal   = {arXiv preprint arXiv:2505.04572},
  year      = {2025}
}

The 37 authors, in the site's order, for checking against a rendered bibliography: Nicolas Hudson, Josh Hooks, Rahul Warrier, Curt Salisbury, Ross Hartley, Kislay Kumar, Bhavana Chandrashekhar, Paul Birkmeyer, Bosch Tang, Matt Frost, Shantanu Thakar, Tony Piaskowy, Petter Nilsson, Josh Petersen, Neel Doshi, Alan Slatter, Ankit Bhatia, Cassie Meeker, Yuechuan Xue, Dylan Cox, Alex Kyriazis, Bai Lou, Nadeem Hasan, Asif Rana, Nikhil Chacko, Ruinian Xu, Siamak Faal, Esi Seraj, Mudit Agrawal, Kevin Jamieson, Alessio Bisagni, Valerie Samzun, Christine Fuller, Alex Keklak, Alex Frenkel, Lillian Ratliff, Aaron Parness.

The BibTeX block above is byte-identical to the entry in the website's Credits section apart from whitespace reflowing of the continuation lines. No author was added, removed or reordered. No @inproceedings entry is given because the site publishes only this @article form.

The original ARMBench paper, for the four Sparrow subsets, is cited separately and is not the right citation for this dataset:

@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}
}

All seven Sparrow authors are present above, in Amazon's order. One character is removed from the entry as published on armbench.com: the stray comma after "Vikedo". BibTeX would otherwise read that field as "last, jr, first" and drop Vikedo Terhuja's given name from the rendered reference. The four Sparrow cards in this suite carry the same correction.

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