Now recruiting capture teams worldwide. Get paid to record first-person video of real work.

Apply to capture

Hard task

Egocentric Chef Knife Work Data

Break down produce and proteins with precise knife cuts and cook dishes on a professional line: that is the catalog’s description of its Hard culinary row, and it is two hands doing different jobs, a blade, and an ingredient that changes shape with every cut. This page covers the egocentric side of one row, Culinary Chef Work / Knife Work in Restaurant & Commercial Kitchen: what first-person chef knife cuts video holds, what the catalog states about the row, and what the site does not state. We run custom kitchen data collection fast and on demand against a spec you write, as an egocentric culinary knife work robot dataset, at scale through a network of thousands of capture providers.
Catalog row
Culinary Chef Work / Knife Work, Restaurant & Commercial Kitchen
Skill and status
Hard, Active collection, as the catalog lists it
Capture setup
Phone (head-mounted) / multiview rig
License
Perpetual, buyer-owned

Why first-person

What does first-person chef knife work video hold that a wide kitchen clip does not?

01

Both hands, the blade and the food in one frame

Knife work is bimanual: one hand holds and guides the ingredient while the other cuts. A head-mounted view keeps both hands, the knife and the produce or protein in frame together, which is the view a robot working at the board would have.

02

A hidden finger is flagged, not guessed

Fingers curl behind an ingredient or a handle. 3D hand pose carries a per-joint visibility flag, so an occluded joint is marked as hidden rather than silently filled in, and contact events mark when a hand meets an item.

03

Steps on one clock

Breaking down an ingredient, cutting and cooking on a line are steps in a chain. Streams are hardware-triggered onto one clock with a cross-stream ceiling under 2 ms, and action segments carry start, end, verb and noun as the skill spec defines them.

04

Masks that follow a changing shape

A cut ingredient is a different object from the whole one. Food-prep captures ship per-frame instance masks rather than masks propagated between frames, so the thing being tracked can change shape between cuts.

Definition

What is egocentric chef knife work data?

It is first-person video of a person doing the work that the catalog row "Culinary Chef Work / Knife Work" describes: breaking down produce and proteins with precise knife cuts and cooking dishes on a professional line. It is recorded from the contributor’s own point of view, so the frames show what the hands are doing to the ingredient and the knife rather than a distant view of someone standing in a kitchen, with synchronized sensor streams alongside.

On this site it is a custom collection. A paid, consented contributor performs the task to a written skill spec. The data is human capture, not robot or teleoperation data, and it is not a fixed corpus you download. People search for the same thing as egocentric knife skills data, first-person chef knife cuts video, a culinary knife work robot dataset or custom kitchen data collection. Those are one request from different angles: name the task, collect it fast and on demand, and receive validated episodes with the labels a training pipeline expects.

The catalog lists this row as Hard. This page is the egocentric treatment of that one row, and the hard task hub links here from its Hard-row table.

Catalog row

Which catalog row does this page rely on, and what is its status?

One. In the commercial catalog, under the Food & Beverage sector and the Restaurant & Commercial Kitchen environment, the row is named "Culinary Chef Work / Knife Work". Its description reads "Break down produce and proteins with precise knife cuts and cook dishes on a professional line." Its skill is Hard, its capture setup is "Phone (head-mounted) / multiview rig" and its status is Active collection. The first table below reproduces it exactly as listed, beside the neighbouring vegetable chopping row.

The environment lists examples such as a sit-down restaurant kitchen, a café, a coffee shop, an artisan bakery and a patisserie. The catalog lists thirteen rows for it, and this is the only Hard one: four are Medium and eight are Easy, as the second table shows. This page names only what the row names: produce, proteins, knife cuts, dishes and a professional line. It does not name a knife type, a cut, an ingredient or a dish, because the row does not.

Active collection is exactly what the catalog states. The site says nothing further about what that means in volume, hours or delivered footage, so this page claims none, and what applies to a given order is confirmed when you request the quote.

Why Hard

Why is chef knife work listed as Hard?

The site gives one definition of Hard: the skill level shown in the catalog, one of Easy, Medium or Hard. It states no criteria beyond that, and this page imports none. What can be said comes from the row text and from copy elsewhere on the site.

The row text asks for two things in one task: precise knife cuts on produce and proteins, and cooking dishes on a professional line. Elsewhere the site describes knife work as exercising fine bimanual dexterity, tool grip and control, and describes the ingredient as changing shape with every cut. One hand guides while the other cuts, so the data has to keep both hands, the blade and the food in view at the same time, and it has to follow an object whose shape is different after each cut.

The site does not state a failure rate, a tolerance, a cut-size standard or a skill level for the people recording, and this page makes none of those claims. Hard here is a catalog label, not a measurement.

Difference

How is this different from the cutting, line cooking and hard task pages?

The egocentric cutting and food prep page is wider. It covers the four catalog rows that name a cutting task, this one among them, together with the documented knife-work category, and describes the whole family. This page is narrower: it covers the single Hard row in the restaurant kitchen environment and what that row states.

The stereo commercial kitchen line cooking page covers this row and the vegetable chopping row from the stereo and MCAP side. It treats a stereo pair as a capture choice agreed in the spec. This page is the egocentric treatment, in the delivery formats the site lists, and it makes no MCAP claim. The egocentric cooking video page covers cooking video in general.

The hard task hub is the index: it lists every row with a skill level of Hard across the catalog and links each one to its task page where one exists. This page is the one that row links to. The comparison table below sets the pages side by side.

Capture setup

What capture setup does the row state, and what is not stated?

The row lists "Phone (head-mounted) / multiview rig", and every one of the thirteen rows in the environment lists the same. This page uses that wording. It does not call the row stereo capture, because the row does not.

The site’s general capture stack is ten streams: a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation and action labels, hardware-aligned under 2 ms. Its instrumented tier is a head-mounted stereo rig, with depth recovered from the calibrated stereo rig. That is the site’s general offer, and the row does not state which tier a knife-work order uses. That is agreed in the skill spec and confirmed on a Spec Pilot. The catalog states that a multiview rig adds synchronized third-person angles when the spec asks for them, and the two wrist cameras give a closer view of the hands than the head camera alone.

The rig is calibrated twice per session, before capture and after it, and an episode is accepted only if RMS reprojection stays under 0.28 px. Depth degrades on wet and reflective surfaces, which is where a rinsed board or a steel counter sits, so kitchen episodes ship per-frame depth confidence maps that let you mask low-confidence regions. Heat is not sensed, and force and tactile readings are not among the streams.

Streams and labels

Which streams and labels come with a knife work episode?

Each attempt is delivered as one episode. Inside it, action segments divide the episode into spans with start_ns, end_ns, verb and noun, and a segment is whatever one step is in the skill spec. The other standard fields are hand_joints and hand_visible for 3D hand pose, contact_events for the moment a hand meets an item, objects, and instance_masks stored at 848×480 so they line up with the depth stream. Hand pose is 21 joints per hand with a per-joint visibility flag, for both hands.

That pairing carries the bimanual structure of the task: where the guiding hand sits relative to the knife hand across a whole step. The site lists no label for the quality of a cut. There is no stated label for cut thickness, uniformity or evenness, no food-state label such as doneness or temperature, no cut taxonomy beyond the verb and noun on each segment, and no force, tactile or thermal stream. If your pipeline needs a vocabulary of steps, the verb and noun list is defined with you in the spec. Anything beyond it is a scoping conversation, not a delivered field. The fourth table below sets each item beside what the site states.

QA

What goes wrong in knife work video, and how is it checked?

The ingredient changing shape is the first risk, because what is being tracked is different after each step. Per-frame instance masks are the site’s answer. Hands are the second: fingers curl behind a piece or a handle, which the visibility flag marks, and a spec can ask for wrist views so a grip the head camera loses is still recorded. Depth on a wet board is the third, handled by confidence maps.

Validation runs against accept criteria written into the spec before capture. Episodes are machine-checked against stated reject criteria: a sync breach beyond the ceiling for more than three consecutive frames, calibration drift beyond tolerance, the manipulated object out of frame for more than 15 percent of the episode, more than 35 percent invalid depth pixels in the manipulation region, an undocumented bystander, or a mismatch against the named skill spec. A human reviewer then checks each episode against the task spec, and a reject log ships with every batch.

Only footage that clears the criteria is billed as a validated hour, and rejected items are recaptured at our cost. Most programs begin with a Spec Pilot, a small first batch that proves the spec and the accept criteria on real footage before volume scales.

Privacy

How are people, items and safety handled?

Every contributor signs a written release before capture, and each delivered episode links to that record through consent artefact IDs. At ingest, automated detection finds faces, plates and screens and blurs them irreversibly. Speech is scrubbed of names, addresses and card numbers while ambient audio is kept, a human reviewer confirms that no un-consented person is identifiable, and footage that cannot be de-identified is discarded. Who may be in frame, and how a kitchen’s own staff, menus or labels are handled, is agreed in the spec.

A contributor knife-handling protocol is not stated on this site, so this page describes none and gives no knife-safety instruction. The site’s stated safety point is about the robot: human capture avoids running an early-stage robot near knives. None of this is legal advice, and the executed agreement controls.

Scale and speed

How fast can a custom chef knife work dataset be delivered?

Custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. That is how the site describes it, and this page adds no provider count, hours or throughput.

Where coverage is deep, the median time from signed spec to first validated batch is a median of 48 hours; where it is thin, turnaround runs three to four weeks. The site does not say which speed applies to a knife work order, so ask when you request the quote, and it states no turnaround for later top-up batches. Delivery is in the format your stack reads: RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or raw streams with metadata. Collections are billed per validated hour, and the buyer owns the dataset under a perpetual, irrevocable licence that is non-exclusive by default, with exclusivity available.

Ordering

How do you order a chef knife work dataset?

Describe what one step is, the produce, proteins and tools and how they begin, what counts as a finished step, and the variation your pipeline is missing. Then name the rig and streams you want, the kitchen types, and how the data should be packaged. That becomes a written skill spec, quoted for reach, timeline and price before any capture begins.

The site’s skill-spec guide asks you to name the verb and the noun rather than a broad activity, to say what geometry must be visible, to set reject rules first and to give one acceptance test. The fifth table below shows what each element decides for a knife work order.

Hard task hub

How does this page fit the hard task hub?

The hard task hub reads the catalog and lists every row with a skill level of Hard: 84 rows across 28 environments, each with its stated status and capture setup. Where a task page exists for a row, the hub links to it, and this page is the link for Culinary Chef Work / Knife Work.

Other Hard rows link to their own pages as those pages are added. A page is linked from the hub only when it states the exact row, so the hub never points at a page that is about something else.

At a glance

What you get.

Catalog row
Culinary Chef Work / Knife Work
Description (as listed)
Break down produce and proteins with precise knife cuts and cook dishes on a professional line.
Environment
Restaurant & Commercial Kitchen, in the Food & Beverage sector
Skill
Hard, the skill level shown in the catalog
Catalog capture setup
Phone (head-mounted) / multiview rig
Status
Active collection
License
Perpetual, irrevocable, buyer-owned, non-exclusive by default

Every collection runs through the same seven-stage end-to-end custom collection pipeline. End-to-end custom data collection is a managed service that takes an AI data need from problem to owned dataset in a single accountable pipeline.

Catalog rows

Which catalog row covers chef knife work?

The Hard row this page relies on, with the neighbouring Medium vegetable row for context. Description, environment, skill, capture setup and status are as listed.
Task (as listed)Description (as listed)EnvironmentSkillCapture setupStatus
Culinary Chef Work / Knife WorkBreak down produce and proteins with precise knife cuts and cook dishes on a professional line.Restaurant & Commercial KitchenHardPhone (head-mounted) / multiview rigActive collection
Chop and portion vegetables using a chef's knife and cutting boardWash, trim and chop vegetables into even pieces and portion them into containers.Restaurant & Commercial KitchenMediumPhone (head-mounted) / multiview rigActive collection

Environment mix

Where does the Hard row sit among the environment’s rows?

All thirteen catalog rows for Restaurant & Commercial Kitchen, counted by skill and status from the catalog. Every row lists the same capture setup.
SkillRowsActive collectionReady to deployCapture setup
Hard110Phone (head-mounted) / multiview rig
Medium431Phone (head-mounted) / multiview rig
Easy853Phone (head-mounted) / multiview rig
All rows1394Phone (head-mounted) / multiview rig

Compared pages

How does this page differ from the neighbouring pages?

Each page, what it covers and how it treats the Culinary Chef Work / Knife Work row.
PageRows it coversTreatment of the knife work row
This page: egocentric chef knife work dataCulinary Chef Work / Knife Work onlyEgocentric treatment of the single Hard row; streams, labels, QA and what is not stated
Egocentric cutting and food prep dataFour cutting rows and the documented knife-work categoryWider cutting family; this row is one of several
Stereo commercial kitchen line cooking dataThis row and the vegetable chopping rowStereo pair and MCAP delivery as a capture and container choice
Egocentric cooking video dataCooking video in generalWider cooking treatment; not specific to this row
Hard task data collectionEvery row with skill HardIndex of Hard rows with status and capture; links here for this row

Capture setup

What does the site state about capture for this row?

Left to right: the item, what the site states, and what that means for a knife work order. “Not stated” means the site does not say.
ItemWhat the site statesFor a knife work order
Catalog capture setupPhone (head-mounted) / multiview rigStated on the row; not called stereo capture
General capture stackTen streams including a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation and action labelsWhich tier a knife work order uses is Not stated; agreed in the spec
Multiview rigAdds synchronized third-person angles when the spec asks for themA spec choice
SynchronizationHardware-triggered onto one clock; aligned under 2 msStated
CalibrationTwice per session; 0.28 px RMS reprojection gateStated
Depth on wet and reflective surfacesDegrades; per-frame confidence maps ship for kitchensStated
Force, tactile and thermal readingsNot among the listed streamsNot stated
Volume, hours and delivered footage for the Active collection rowNot statedConfirmed when you request the quote

Streams and labels

Which streams and labels does a knife work episode carry?

Delivered items are stated elsewhere on this site. Items marked not stated are not promised on this page.
Stream or labelWhat it gives a knife work episodeStatus
Head and wrist camerasFirst-person view of the work plus a closer view of the handsDelivered
Depth with per-frame confidence mapsBoard and ingredient geometry; low-confidence regions can be maskedDelivered
3D hand pose, 21 joints per hand, with visibility flagsBoth hands across a step; occluded fingers flaggedDelivered
Contact eventsThe moment a hand meets an itemDelivered
Action segments (start_ns, end_ns, verb, noun)One labeled span per step as the spec defines itDelivered
Per-frame instance masks (848×480)Ingredient masks that follow a changing shapeDelivered
Consent artefact IDs and anonymization recordProvenance per episodeDelivered
Cut-quality, thickness or uniformity labelNot listedNot stated
Food-state label (doneness, temperature)Not listedNot stated
Force, tactile or thermal readingsNot among the listed streamsNot stated
Robot or teleoperation dataHuman first-person demonstrations onlyNot captured

Skill spec

What does the skill spec fix for a chef knife work order?

Spec element and where a buyer can see it in the delivery. “Set per program” means the site states no fixed value.
Spec elementWhat it decidesWhere it shows up
Named skillThe verb and the noun for each step, not a broad activityAction segment labels; set per program
Geometry that must be visibleWhat must be in frame and for how long: both hands, the blade, the foodAccept criteria and reject log
Ingredients, tools and start stateWhich produce, proteins and tools, and how they beginSet per program; the site states no cut or doneness label
Episode boundaryOne step or a chain of steps on a lineSet per program; the site does not state an episode length
Rig and streamsHead camera, wrist views, multiview angles, hand pose, depth, IMUThe channels you can read; confirm on the Spec Pilot
Reject rule and acceptance testWhat fails an item and what counts as completeReject log per batch
Delivery formatHow the data is packagedRLDS, LeRobot, WebDataset, HDF5, zarr, Rerun .rrd or raw streams
PrivacyHow people and items in frame are protectedFaces, plates and screens blurred at ingest
LicenseWho owns the datasetPerpetual, irrevocable, buyer-owned, non-exclusive by default

Example footage

What does egocentric kitchen footage look like?

Two real clips already published on this site. The first is a home-kitchen clip of an operator browning mince at a gas range, listed in the site’s Kitchen & food prep set. The second is a format and rig example: a head-mounted stereo capture with hands in frame. Neither shows a professional line or chef knife work.

sample-dinner-cooking.mp4 · Kitchen & food prep · home kitchen, stirring in a skillet · 640 × 360 web preview · not a professional line or knife-work clip
egocentric_stereo_sample.mp4 · head-mounted stereo · hands in frame · 1920 × 1080 web preview · format and rig example

No chef knife work clip is published on the site, so these are labelled as format and rig examples. Request a sample to see delivered files. See sample data outputs, stereo sample page, request an egocentric data sample.

FAQ

Egocentric knife skills data, answered.

What is egocentric chef knife work data?
Custom first-person video of a person breaking down produce and proteins with precise knife cuts and cooking dishes on a professional line, with hand pose and sensor streams alongside, collected on demand to a written spec.
Which catalog row covers it, and what is its status?
Culinary Chef Work / Knife Work in Restaurant & Commercial Kitchen: skill Hard, capture setup Phone (head-mounted) / multiview rig, status Active collection. This page claims no volume, hours or delivered footage.
What does Hard mean here?
Only the skill level shown in the catalog, one of Easy, Medium or Hard. The site states no further criteria, and this page gives no failure rate, tolerance or skill claim about workers.
Is this stereo capture?
Not stated on the row. The catalog lists a head-mounted phone or multiview rig. A stereo pair is part of the site’s general stereo offer and is agreed in the skill spec; the stereo commercial kitchen line cooking page covers that side.
Does it include cut-quality, force or food-state labels?
Not stated. The delivered fields are action segments, 3D hand pose with visibility flags, contact events, objects and per-frame instance masks, plus depth confidence maps. No cut-quality, thickness, force, tactile, thermal or food-state data is listed.
How is the data checked before delivery?
Episodes are machine-checked against reject criteria written into the spec, such as the object out of frame for more than 15 percent of the episode or more than 35 percent invalid depth pixels, then reviewed by a human. A reject log ships with each batch.
How are faces and screens handled?
Faces, plates and screens are detected and irreversibly blurred at ingest with reviewer confirmation, and footage that cannot be de-identified is discarded. The executed agreement controls, and this is not legal advice.
How fast is a first batch, and who owns the data?
Where coverage is deep, the median time from signed spec to first validated batch is a median of 48 hours; thin coverage runs three to four weeks. The buyer owns the dataset under a perpetual, irrevocable, non-exclusive-by-default licence, with exclusivity available.

Sources

Where are the formats and the public kitchen dataset documented?

Request a custom chef knife work dataset.

Tell us what one step is, the produce, proteins and tools and their start state, the kitchen types, and the streams and format you want. We will quote reach, timeline and price against a written spec.