amazon/armbench-stow-success-prediction

ARMBench - Stow Success Prediction

This repository is the official Amazon redistribution, on the Hugging Face Hub, of the Stow Success Prediction subset of ARMBench. The archive under archive/ is Amazon's original release object, byte-for-byte: a single ZIP64 archive holding 290,608 JPEG images and one Parquet feature table.

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 paper. This subset comes from the Vulcan Stow program and is cited with hudson2025stow (arXiv:2505.04572), not with the Sparrow ARMBench paper (arXiv:2303.16382). See Citation.

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 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:

Stow Success Prediction: Complete dataset of 72,000+ stow cycles with bin images, item images, and tabular data. Predict binary outcomes (success/failure) and space creation in fabric-covered pods.

The task, verbatim from the Stow Success Prediction page

The Stow Success Prediction task focuses on predicting the binary outcome (success or failure) of a stow attempt in Amazon Vulcan Stow workcells. In the robotic manipulation context, this task is critical for optimizing warehouse operations and improving the efficiency of robotic stowing systems.

The Stow Success Prediction dataset contains 72,000+ complete stow cycles with data collected from Amazon Vulcan Stow workcells located in an Amazon warehouse. The dataset supports the Success Prediction task: Predicting the binary outcome of a stow attempt (success or failure) based on the provided images and tabular data. The dataset is organized as follows:

Images: All images are named after their corresponding run_id for cross-referencing:

• Induct: Items on the conveyor before stowing
• Pre-stow Pedestal: Bin before stowing
• Mid-stow Pedestal: Bin during the stow process
• Post-stow Pedestal: Bin after stowing

The pedestal images are captured from fixed cameras, showing the specific bin being stowed into at different stages of the stow process.

Tabular Data: A comprehensive table with run_id as the primary identifier and all other feature columns as described below.

Editorial note on the quotations above, so the differences from the live page are auditable and not mistaken for transcription errors: bold and list markup were applied to the page's own headings and bullets, and the four image stages were rendered as a list because the page renders them as one. No word was changed, added, reordered or removed.

One point of interpretation, clearly marked as not Amazon's words. The home-page blurb names two prediction targets, "binary outcomes (success/failure) and space creation", but the task page defines only the binary one. The archive ships no separate space-creation label; the columns the site groups under Space Metrics (mlo, dlo, dao, dco, dloSweep) are the natural regression targets for that second task. That mapping is this card's reading, not a statement published by Amazon.

Dataset Structure

What this repository contains

Path	Bytes	Description
archive/success_prediction.zip	40,166,726,870	Amazon's original release archive, unmodified

Apart from this card (README.md), that is the whole repository. 40,166,726,870 bytes is 37.408 GiB.

Amazon publishes no checksum sidecar for this subset - unlike the Sparrow defect archives, there is no .sha256 file next to the release object, and no -tiny sample archive exists for Stow Success Prediction. See Verifying what you downloaded for what can be checked instead.

There is no dataset viewer for this repository, deliberately. The Parquet feature table lives inside the ZIP, and a configs: block pointing at a path inside an archive does not resolve - it would break the viewer rather than populate it. The recipe below pulls the table out without downloading all 37.4 GiB.

Archive contents

Measured from the archive's own ZIP64 central directory (not estimated):

Format	ZIP64, deflate (method 8) on every member sampled
Total central-directory entries	290,615
of which JPEG images	290,608
of which Parquet tables	1
of which directory entries	6
Central directory	35,784,783 bytes (34.127 MiB) at offset 40,130,941,989

Layout:

image/
  induct/<run_id>.jpg                 72,652 files
  primary_pedestal_pre/<run_id>.jpg   72,652 files
  primary_pedestal_mid/<run_id>.jpg   72,652 files
  primary_pedestal_post/<run_id>.jpg  72,652 files
tabular/
  tabular_data.parquet                72,652 rows

Member group	Files	Uncompressed bytes	Size	Share of image bytes	Mean file	Min	Max
image/induct/	72,652	20,972,795,900	19.532 GiB	52.4%	288,675	245,018	605,497
image/primary_pedestal_pre/	72,652	6,468,060,192	6.024 GiB	16.1%	89,028	2,101	297,016
image/primary_pedestal_mid/	72,652	6,082,493,287	5.665 GiB	15.2%	83,721	2,011	299,787
image/primary_pedestal_post/	72,652	6,530,314,460	6.082 GiB	16.3%	89,885	1,991	295,825
all images	290,608	40,053,663,839	37.303 GiB	100%	137,827	1,991	605,497
tabular/tabular_data.parquet	1	106,432,610	101.502 MiB	-	-	-	-

The archive is 6,630,421 bytes larger than the sum of its uncompressed members (40,166,726,870 vs 40,160,096,449). That is expected: deflate gains essentially nothing on already-encoded JPEG, while the container adds 290,615 local file headers and a 34.127 MiB central directory. Do not read the ZIP as a compressed archive - read it as a seekable container.

Image resolutions, from a single sampled cycle (run_66e2d40aea, one image per stage) and therefore not asserted for the whole release: induct 1224x1024, and all three pedestal stages 521x273. The per-stage min/max byte spreads above (pedestal images range from 1,991 to ~300,000 bytes) make it likely that resolution or content varies across cycles; Amazon's site does not publish image dimensions, and no dimension survey was run over all 290,608 files.

The run_id join, verified

The site says the images are "named after their corresponding run_id for cross-referencing". That holds exactly, and it was checked rather than assumed. Over the full central directory and the full Parquet table:

• run_id in the Parquet table: 72,652 rows, 72,652 distinct, zero nulls.
• Each of the four image stages contains exactly 72,652 files, and each stage's set of filename stems is equal to the Parquet's run_id set - not a subset, not a superset, for all four stages.
• Therefore the set of run_ids with all four images present is all 72,652 of them.

So one stow cycle is exactly one Parquet row plus exactly four JPEGs, with no orphans in either direction.

The tabular table

Rows (stow cycles)	72,652
Top-level fields	236
Physical leaf columns	240
Row groups	1
created_by	parquet-cpp-arrow version 20.0.0
Size inside the ZIP	95,166,029 bytes deflated / 106,432,610 bytes uncompressed

The 236-versus-240 gap is not a discrepancy: four fields are nested (...etiquetteViolations list<string>, ...pBinToProfileBottom and ...pBinToProfileTop list<struct<x,y,z>>, ...binHSweepWaypoints list<null>), and Parquet counts their leaves, so 236 - 4 + 8 = 240.

The single row group means the table is read in one pass; it cannot be streamed by row group.

The release ships no official train/test split. There is no split column, no train.csv/test.csv, and the website publishes no split for this task. Any split is yours to define - say so in your paper, and prefer a split that does not leak, since consecutive cycles can share a bin.

Column schema, verbatim from the site

The following table is Amazon's published schema, reproduced with the site's own category grouping, feature names, descriptions and data types. It documents 42 features (including run_id).

Category	Feature Name	Description	Data Type
Metadata	run_id	Unique identifier for each stow cycle	string
Outcome Features	computed_success	Boolean indicating if the stow attempt was a success (True) or failure (False)	bool
	computed_unproductive	Boolean indicating if the stow attempt was unproductive (True) or not (False)	bool
	computed_defect	Boolean indicating if the stow attempt was a defect (True) or not (False)	bool
Item Features	itemLengthPerception	Item length estimate produced by perception [m]	float
	itemWidthPerception	Item width estimate produced by perception [m]	float
	itemHeightPerception	Item height estimate produced by perception [m]	float
	itemLengthManifest	Item length estimate from the manifest [m]	float
	itemWidthManifest	Item width estimate from the manifest [m]	float
	itemHeightManifest	Item height estimate from the manifest [m]	float
	weightKg	Weight of item in KG (from manifest)	float
	itemPackagePackagingType	String indicating the item packaging type (box, polybag, etc.)	string
	itemPackageShape	String indicating the item packaging shape	string
	isFragile	Boolean for if the item is fragile	bool
	isFoldable	Boolean for if the item is foldable	bool
	containsLiquidClassification	Boolean for whether the item contains liquid	bool
	bindingText	String indicating the use category of the item (e.g., 'Kitchen', 'Office Product', 'Apparel')	string
Bin Features	binDepth	Bin depth measurement [m]	float
	binWidth	Bin width measurement [m]	float
	binHeight	Bin height measurement [m]	float
	binVolume	Bin volume measurement [m³]	float
	binGauPerception	Perception estimate of the gross area utilization [m²]	float
	gluScore	Metric for gross linear utilization	float
	gcuScore	Metric for gross cubic utilization	float
	numItems	Number of items in the bin prior to the current stow attempt	int
	numStows	Number of stow attempts to this bin prior to the current attempt	int
	numSuccessfulStows	Number of successful stows to the bin we are attempting to stow to	int
	isEdge	Boolean indicating if the bin we are attempting is on the edge of the pod or not	bool
	podType	Categorical feature indicating the type of pod being stowed to	string
Stow Strategy Features	stowStrategy.type	String indicating the stow strategy type (e.g., SLOT_WEDGE)	string
	stowStrategy.direction	Blade sweep direction ('MINUS_Y', 'PLUS_Y')	string
	stowStrategy.slotEdgeType	String for slot edge type ('NO_EDGE', 'RIGHT', 'LEFT')	string
	ssBinSweep	Boolean indicating if stow attempt was a bin sweep	bool
	ssSlotWedge	Boolean indicating if stow attempt was a slot wedge	bool
	ssStacking	Boolean indicating if stow attempt was stacking (rotated wedge)	bool
	slotWidth	Width of the slot we are stowing to in this cycle [m]	float
	maxL3SlotWidth	Maximum slot width of all BinSweep affordances (L3 type slots)	float
Space Metrics	mlo	Missed linear opportunity metric	float
	dlo	Detected linear opportunity metric	float
	dao	Detected area opportunity metric	float
	dco	Detected volume opportunity metric	float
	dloSweep	Detected linear opportunity metric for Bin Sweep (L3) affordances	float

How those names map onto the Parquet file

The Parquet column names are not the bare names in the table above. They are namespaced, and the feature block appears twice. Nothing is missing except podType. The mapping, measured from the file:

Site's name	Parquet column(s)
run_id, computed_success, computed_unproductive, computed_defect	identical, top level
Item dimensions, weightKg, all Bin Features except podType, all Stow Strategy Features except the three stowStrategy.*, all Space Metrics	match_risk_features.<name> and match_risk_features.initial.<name>
itemPackagePackagingType, itemPackageShape, isFragile, isFoldable, containsLiquidClassification, bindingText	semantic_summary.skumasterAttr.<name>
stowStrategy.type, stowStrategy.direction, stowStrategy.slotEdgeType	match_output.stowStrategy.<name>
podType	not present in the Parquet file

The two feature snapshots are the single most important structural fact for anyone training on this table:

• match_risk_features.initial.* - 64 fields
• match_risk_features.* (non-initial) - 64 fields, the same 64 names

The site does not document the difference between them. Treat them as two snapshots of the same feature vector and do not feed both into a model without checking which one is causally available before the stow decision, or you risk leaking outcome information.

Field counts by namespace, measured: match_risk_features.initial.* 64, match_risk_features.* 64, semantic_summary.* 20, match_output.* 84, plus run_id and the three computed_* outcomes = 236.

The site documents 42 features; the file carries 236 top-level fields. The ~194 undocumented ones include full 6-DoF blade and item insert poses (match_output.stowStrategy.binHBladeInsert.*, ...binHItemInsert.*, ...worldHBin.* as translation + quaternion), predicted slot corner coordinates, band-manipulation waypoints, item-in-hand overhang offsets in cm, insertion costs, the additional space metrics mao/mco/tao/tco/tlo/dloFlip/dloL2/dloStack, and ten semantic_summary.operatorAnnotation.* fields whose names indicate human operator entries (isStowedItemDamage, isPodDamage, bandsNotClosed, etiquetteViolations, ...). Amazon's website does not document any of these, and this card does not guess at their semantics beyond what the names say.

Label distribution, measured over all 72,652 rows

Outcome column	True	False	True rate
computed_success	62,433	10,219	85.935%
computed_unproductive	6,551	66,101	9.017%
computed_defect	3,667	68,985	5.047%

The headline task is therefore an imbalanced binary problem with a 14.065% failure rate (10,219 of 72,652). Report precision/recall or average precision on the failure class; accuracy is dominated by the majority class. The three outcome columns are independent booleans, not one 3-way label.

Measured value sets for the categorical columns, as published in the file (strings are carried with their enum prefixes intact - strip them if you want readable labels):

Parquet column	Distinct	Values and counts
match_output.stowStrategy.type	3	SLOT_WEDGE 55,942; BIN_SWEEP 15,688; STACKING 1,022
match_output.stowStrategy.direction	2	MINUS_Y 68,021; PLUS_Y 4,631
match_output.stowStrategy.slotEdgeType	3	NO_EDGE 48,356; LEFT 14,121; RIGHT 10,175
semantic_summary.skumasterAttr.itemPackagePackagingType	7	ItemPackagePackagingType_SINGLE_ITEM_MINIMAL 34,910; ..._NONE_ITEM_PACKAGE_TYPE 19,590; ItemPackagePackagingType_ (empty suffix) 13,369; ..._MULTI_ITEM_MINIMAL 3,967; ..._PAPER_BAG_PAPER_ENVELOPE 552; 2 more
semantic_summary.skumasterAttr.itemPackageShape	5	ItemPackageShape_CAN_ROLL 33,099; ItemPackageShape_ (empty suffix) 23,118; ..._OTHER_SINGLE_SHAPE 16,223; ..._NONE_ITEM_PACKAGE_SHAPE 211; 1 null
match_risk_features.isEdge	2	'0' 52,568; '1' 20,084
semantic_summary.skumasterAttr.bindingText	61	'' (empty) 13,048; Health and Beauty 8,744; Apparel 7,716; Kitchen 6,331; Unknown Binding 5,415; 56 more

Known data quirks

Measured properties of this release, documented so you do not have to rediscover them.

1. isEdge is a string, not a bool. The values are the characters '0' and '1'. The site's schema calls it bool. Cast explicitly.
2. numItems, numStows, numSuccessfulStows, numberFlat, numberInstances, stowPrevious are strings, not ints. The site's schema calls the first three int. numItems has 35 distinct string values. Cast explicitly, and decide what to do with any non-numeric value before you do.
3. podType does not exist in the file. The site documents it; a case-insensitive search for pod over all 236 field names returns only semantic_summary.operatorAnnotation.isPodDamage.
4. Strategy-specific blocks are null exactly where the strategy does not apply, and the counts close perfectly. match_output.stowStrategy.slotWedge.* is null in 16,710 rows, and 72,652 - 16,710 = 55,942 = the SLOT_WEDGE count. ...binSweep.* is null in 56,964 rows, and 72,652 - 56,964 = 15,688 = the BIN_SWEEP count. These nulls are structural, not missing data.
5. Exactly one row has a missing semantic summary. All 20 semantic_summary.* fields have a null count of exactly 1 - the same row. No other field outside the strategy blocks in quirk 4 has any null.
6. Categorical strings carry their enum prefix, and some carry an empty suffix: ItemPackagePackagingType_ (13,369 rows) and ItemPackageShape_ (23,118 rows) are "present but unspecified", distinct from both null and from NONE_ITEM_PACKAGE_TYPE. Three states, not two.
7. The feature vector is duplicated under match_risk_features.* and match_risk_features.initial.*. See the leakage warning above.
8. The directory names are not the site's stage names. primary_pedestal_pre / _mid / _post are the site's "Pre-stow / Mid-stow / Post-stow Pedestal". The primary_ prefix suggests a non-primary camera exists in the source system; none is in this release.

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
Stow cycles	72,000+ (home page and task page)	72,652 (Parquet rows, distinct run_ids, and files per image stage - all three agree)
Images	not published as a count	290,608 JPEGs = 4 x 72,652
Download size	not published for this subset	40,166,726,870 bytes = 37.408 GiB
Documented tabular features	42 (including run_id)	41 of 42 present; podType absent; 236 top-level fields in total
numItems, numStows, numSuccessfulStows	int	string
isEdge	bool	string '0'/'1'
Column names	bare, e.g. itemLengthPerception	namespaced, e.g. match_risk_features.itemLengthPerception, with a .initial. twin
Image stage names	Induct / Pre-stow / Mid-stow / Post-stow Pedestal	image/induct/, image/primary_pedestal_pre/, _mid/, _post/

The 72,000+ versus 72,652 relationship is consistent - "72,000+" is a rounded-down published figure, not a conflicting measurement. The type and naming rows above are genuine divergences between the published schema and the shipped file; the file is authoritative for what you will load, and the site is authoritative for what each feature means.

Verifying what you downloaded

Amazon publishes no SHA-256 for this subset, so there is no official checksum to check against. What can be verified:

hf download amazon/armbench-stow-success-prediction \
  archive/success_prediction.zip \
  --repo-type dataset --local-dir ./armbench-stow

# 1. exact size
stat -c %s ./armbench-stow/archive/success_prediction.zip    # 40166726870

# 2. structural integrity and the full member list (reads the central directory only)
unzip -l ./armbench-stow/archive/success_prediction.zip | tail -3
#   expect 290615 entries, of which 290608 *.jpg and 1 *.parquet

# 3. per-stage counts
unzip -l ./armbench-stow/archive/success_prediction.zip \
  | grep -oE 'image/[a-z_]+/' | sort | uniq -c
#   expect 72652 for each of induct, primary_pedestal_pre, _mid, _post

For provenance: the source S3 object's ETag is 050900df7ed32ddff192a71cf2d4bbc5-767. The -767 suffix means it is a 767-part multipart ETag, not an MD5 of the file, so it cannot be reproduced with md5sum. It is quoted only so the release object can be matched against the source bucket.

Getting the feature table without downloading 37.4 GiB

With the whole archive in hand:

unzip -p success_prediction.zip tabular/tabular_data.parquet > tabular_data.parquet
python -c "import pyarrow.parquet as pq; t=pq.read_table('tabular_data.parquet'); print(t.num_rows, len(t.schema))"
# 72652 236

A ZIP is a seekable container, so a Range-capable HTTP client can fetch only the table. The byte offsets below were measured from this exact archive and are stable for it:

Local file header of tabular/tabular_data.parquet	offset 40,035,775,874
Deflated member length	95,166,029 bytes
Uncompressed length	106,432,610 bytes
Compression	deflate (method 8), raw stream, zlib.decompress(blob, -15)

Fetch bytes=40035775874-40131000000, skip the 30-byte local header plus its name and extra fields, then inflate. The same two-step trick reads the central directory first: the ZIP64 end-of-central-directory record is in the last 1 KiB, and the central directory itself is 35,784,783 bytes at offset 40,130,941,989. This was the method used to measure every figure in this card, and it transfers about 131 MB instead of 37.4 GiB. The offsets are a property of the archive, not of the Hub; whether a given client honours Range requests against a Hub resolve URL has not been tested here.

Dataset Creation

The data was collected from Amazon Vulcan Stow workcells in an Amazon warehouse, as stated on the task page and quoted above: "72,000+ complete stow cycles with data collected from Amazon Vulcan Stow workcells located in an Amazon warehouse."

The outcome columns are named computed_*, which indicates they were derived by the production system rather than hand-labelled; the semantic_summary.operatorAnnotation.* fields indicate human operator entries. Amazon's website does not publish the annotation protocol, the definition used to compute success, unproductive or defect, the annotator population, or inter-annotator agreement, and this redistribution does not add them. If the precise definition of computed_success matters to your result, treat it as unpublished and say so.

The source S3 object is dated 2025-07-23, consistent with the August 2025 release entry below.

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.

Considerations for Using the Data

Intended use. Research and benchmarking of stow outcome prediction - predicting, from bin and item images plus tabular features, whether a robotic stow attempt into a fabric-covered pod will succeed.

Class imbalance. 14.065% of cycles are failures. Report per-class precision/recall or average precision on the failure class.

No official split. Results on this dataset are only comparable if the split is stated. There is none in the release, so publish yours.

Leakage risks, two of them. First, the duplicated match_risk_features.* / match_risk_features.initial.* blocks: the site does not say which snapshot is taken before the stow decision, so using both without checking may leak. Second, the images: primary_pedestal_mid and primary_pedestal_post are captured during and after the stow, so a model that sees them is not solving the prediction problem the task page describes - it is observing the outcome. A genuine success predictor has access to induct and primary_pedestal_pre only. The site does not state this restriction; it follows from its own stage definitions, quoted above.

Scope limits. The data comes from Vulcan Stow workcells in one Amazon warehouse, under that facility's lighting, camera placement, pod geometry and inventory mix. Models trained here should not be assumed to transfer to other workcells, pod types or inventory distributions without re-evaluation.

Personal and sensitive information. The images show items on a conveyor and inside fabric pods. 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, contact armbench-support@amazon.com.

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. archive/success_prediction.zip is byte-for-byte identical to Amazon's release object, 40,166,726,870 bytes.
Additions	None. This repository contains no derived files. Every figure in this card was measured from the archive; none of it was added to the archive.
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 Vulcan Stow paper, not the Sparrow ARMBench paper.

The ARMBench suite on the Hub

ARMBench is distributed as six downloads, each published as its own dataset under the amazon organization so 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 (this dataset)	Vulcan Stow	Stow success prediction	37.4 GiB
amazon/armbench-blade-insertion-affordance	Vulcan Stow	Blade insertion affordance	555.7 GiB

The four Sparrow subsets are cited with mitash2023armbench; the two Vulcan Stow subsets, including this one, are cited with hudson2025stow.

Citation

Amazon's published BibTeX entry for the Vulcan Stow datasets, reproduced from the stow-citation-text block on armbench.com. All 37 authors are present, in the site's order.

@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 only change from the site's text is the removal of trailing whitespace at the end of the wrapped author lines. No author was added, removed, reordered, abbreviated or replaced by "and others"; dropping an author would be an attribution failure under CC BY 4.0. The 37 authors, in order, are: 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.

If you also use a Sparrow subset, cite mitash2023armbench for that subset as well; see amazon/armbench-defect-video for that entry.

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