One named catalog row
The task is the catalog row "Stir + flip on stovetop", described as "Stir a pan and flip contents." A spec names that row rather than relying on a generic cooking label.
Now recruiting capture teams worldwide. Get paid to record first-person video of real work.
Apply to captureTask row

Requirements
The task is the catalog row "Stir + flip on stovetop", described as "Stir a pan and flip contents." A spec names that row rather than relying on a generic cooking label.
Stirring and flipping is two-handed work close to the pan. Hand pose carries a per-joint visibility flag, so a finger hidden behind an object is flagged rather than guessed.
The row words two actions in one description: stir a pan and flip contents. The spec names which you need, in what order, and where each starts and ends, so each becomes its own labeled segment.
No recipe, contents, pan or utensil type, heat setting, task time or quality claim is made, because the site states none for this row.
Definition
Egocentric stovetop stirring and flipping data is first-person video of a person stirring a pan and flipping what is in it, recorded from their own point of view with synchronized sensor streams alongside. The task wording comes from one catalog row, "Stir + flip on stovetop", which the catalog describes as "Stir a pan and flip contents." This page repeats that description as written and adds nothing to it. The data is human capture: a paid, consented contributor performs the task to a written skill spec. It is not robot or teleoperation data, and it is not a fixed corpus you download.
People reach this topic through several phrasings: egocentric stovetop cooking data, first-person pan stirring and flipping video, a handheld gripper pan-flip dataset, or custom cooking data collection. They describe one request from different angles. Under each of them the work is the same: write down the task, collect it on demand, and receive validated episodes with the labels a training pipeline expects. Nothing on this page states that the video carries a contents-state label, a stir-progress label or a flip-success label.
The task row
One row names the task. In the task catalog, the Working Kitchen Set (food prep) environment, in the Food & Beverage sector, lists "Stir + flip on stovetop" with the description "Stir a pan and flip contents." It is rated Medium, its capture mode is Multiview rig + handheld grippers, and its status is Available on request, meaning the work is scoped and recruited to spec when a buyer asks.
The catalog gives one line for the row. What is in the pan, how many stirs or flips an attempt holds and where an attempt starts and ends are set by your spec. The first table below gives the row exactly as the catalog states it. This page claims no delivered hours for it.
The setting
The catalog describes the environment with these examples: stovetop, counter, sink, dishwasher, utensils and food props. Fourteen rows sit in it, and every one lists the same capture mode, a multiview rig with handheld grippers, and the same status, Available on request. Ten are rated Medium and four are rated Easy.
The site describes this capture mode as a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers, so demonstrations map onto a robot end effector. The site states that Firsthand does not capture at a customer’s own facility, so the stovetop is a capture setting arranged for the task, not your kitchen.
Nearby rows
Several neighbouring rows in the same environment share parts of a cooking job. "Crack + whisk eggs" reads "Crack eggs and whisk, then pour to pan." "Measure + mix" reads "Measure ingredients and mix a batter." "Knead + roll dough" reads "Knead dough and roll it flat." "Assemble a salad" reads "Tear greens, add toppings, dress, and toss a salad." "Peel + chop vegetables" reads "Peel and chop vegetables with a knife." All of them are rated Medium, listed with Multiview rig + handheld grippers and Available on request.
These are separate rows, not steps of the stovetop row. They are context only, and each is its own request. If your policy needs the egg step that ends in a pour to a pan, or the chopping that comes before the pan, that is a spec question for the neighbouring row.
Related pages
The cooking page is organized around a domain: first-person food-prep and cooking capture, with the kitchen task categories the site documents. Stovetop transfer with a loaded pan is one of those categories there, so stovetop work is one category among several rather than the main topic. The egg page keeps a single row, Crack + whisk eggs, which ends with a pour to a pan. The baking page covers two dough and batter rows. The stereo cutting page covers peeling and chopping vegetables with a knife.
This page keeps a single row, Stir + flip on stovetop, and looks at what a capture of it needs. The comparison table below sets the pages side by side. For cooking as a whole, start from the cooking page. For eggs, dough or knife work, start from the matching page. For stirring a pan and flipping its contents, start here.
Why it is hard
The row bundles two actions in one description: stir a pan and flip contents. A spec has to say which of those it needs, in what order, and where each one starts and ends, so each can become its own labeled segment. The site states no recipe, contents, pan or utensil type, stirring pattern, flip count or duration, so none is claimed here.
Limits the site already states apply. Hand occlusion is the first: fingers disappear behind objects in manipulation, which is why 3D hand pose carries a per-joint visibility flag. Depth is the second: it degrades on wet and reflective surfaces, and per-frame confidence maps mask the low-confidence regions rather than hiding them. The cooking page adds that steam and heat degrade sensors and that ingredients change shape and state, and that masks propagated frame to frame do not track them well, so cooking captures ship per-frame instance masks. How the contents of a pan behave in depth is not stated for this row. If your policy depends on it, say so at the start.
An episode is also rejected if the manipulated object is out of frame for more than 15 percent of it, so the spec decides what the manipulated object is: the pan, the utensil or the contents.
Capture setup
The catalog lists one capture mode for the row: a multiview rig with handheld grippers. The site describes that mode as a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers. The site does not explain the grippers further and does not state a camera count for the row.
The instrumented tier is a head-mounted stereo rig worn by the contributor, with first-person RGB and metric depth recovered from the calibrated stereo rig. The wider stack is ten streams: a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation and action labels, hardware-triggered onto one PTP clock domain and aligned under 2 ms, with 6-DoF IMU data at 200 Hz. The rig is calibrated twice per session, and an episode is accepted only if RMS reprojection stays under 0.28 px. Which tier applies to this row is not stated and is confirmed in the skill spec.
This is an egocentric page. The catalog does not list stereo capture for the Stir + flip on stovetop row, and this page does not describe the task as stereo capture. Stereo applies only through the site’s general stereo offer, agreed in the skill spec.
Handheld grippers
The catalog lists handheld grippers because the operator holds a gripper rather than a bare hand while performing the task, and the site says this lets demonstrations map onto a robot end effector. For a stir-and-flip row that matters because the motion of the utensil and the pan is the thing a policy has to reproduce. A handheld gripper pan-flip dataset keeps the demonstration closer to the hardware a robot would use than a bare-hand clip does.
The site does not state which gripper is used, how it attaches or what its dimensions are, and this page names none. It also does not claim that a model trained on this capture succeeds at the task. The embodiment mapping is a design choice agreed in the skill spec, and the site does not claim that human capture replaces robot data.
Episodes
Each attempt is delivered as one episode. The site does not set an episode length, so the skill spec decides what one attempt is: a single stir, a single flip, or the whole run from first stir to the last flip. Inside the episode, the synchronized streams run from start to end, and action segments divide it into labeled spans. Whether the stir step and the flip step stand alone, as they are worded together in this row, or sit inside a longer cooking sequence is a spec decision.
The standard fields are action segments with start_ns, end_ns, verb and noun; hand_joints and hand_visible for 3D hand pose on both hands, with a per-joint visibility flag; contact_events for the moment a hand meets an object; an objects list; and per-frame instance masks stored at the depth resolution of 848×480. The site does not say whether a pan, a utensil or its contents counts as the item in a contact event, so the spec settles it.
Labels and delivery
Validation runs against signed accept criteria written into the spec before capture, and a reject log ships with each batch. Episodes are machine-checked against reject criteria, including 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 and a mismatch against the skill spec. A human reviewer then checks each episode. Only footage that clears the criteria is billed as a validated hour, and rejected items are recaptured at our cost.
Delivery is in the format your stack already reads: RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, or raw streams with metadata, with readable converter source alongside. The license is perpetual, irrevocable and worldwide, buyer-owned and non-exclusive by default, with exclusivity available as a separate term. The executed agreement controls.
Capture and privacy
Every contributor signs a written release, and each delivered episode links to it through consent artefact IDs in the batch data card. At ingest, automated detection finds faces, plates and screens and blurs them irreversibly. Speech is scrubbed of names, addresses and card numbers. A human reviewer confirms that no un-consented person is identifiable in any frame, footage that cannot be de-identified is discarded, and the anonymization record ships with the data.
A kitchen can put printed text in view, such as a label on a package. Whether a product label is blurred is not stated, so it is agreed in the spec. A kitchen can have other people nearby, and an undocumented bystander is a reject criterion. The site states no contributor safety protocol and no food-safety guidance for any kitchen row, and this page does not describe one.
Not stated
Where this site does not state something, this page says "not stated" rather than filling the gap. There is no listed label for contents state, stir progress, flip success, recipe or quantity, no force or tactile stream, no thermal stream, and no stated cycle time, error rate or success rate for the row. The site states no heat setting, temperature or cooking time for any kitchen row. Counts of contributors, episodes and hours are set in your skill spec. Nothing here claims a training outcome or a model improvement.
The tradeoff to plan for is the embodiment gap between a person stirring and flipping with a handheld gripper and a robot platform doing it, which retargeting or pairing with robot data has to close. No named model was trained on this capture, and this page makes no claim about any.
Custom collection
The site lists an egocentric kitchen video dataset as a fixed corpus you can review, but the row for this task is Available on request, so a stovetop stirring and flipping collection is custom collection. It runs as custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. This page does not state a provider count, hours or throughput beyond what the site already states.
Because the row is Available on request, the work is scoped and recruited to your spec rather than pulled from a library of existing hours. If you also need the egg, dough or chopping steps around the pan, the neighbouring rows are listed with their own statuses in the first table.
Speed
From signed spec to first validated batch, the published timeline is a median of 48 hours where coverage is deep and three to four weeks where it is thin and contributors need to be recruited first. The site does not say which applies to the working kitchen set, so the range that applies is confirmed when the spec is quoted. At program level, a pilot runs two to three weeks to first delivery and a standing program reaches full rate in four to six weeks.
Getting started
Describe the steps you need, their order, where an attempt starts and ends, what counts as complete and what the camera must see. That becomes a written skill spec, quoted for reach, timeline and price before capture. Most programs begin with a Spec Pilot, a small, fast first batch that proves the spec before volume scales. Collections are billed per validated hour.
Capture setup
| Part | As stated on this site |
|---|---|
| Catalog capture mode | Multiview rig + handheld grippers: a staged working set recorded by a multiview rig while the operator performs the task with handheld grippers |
| Instrumented tier | Head-mounted stereo rig worn by the contributor; first-person RGB with metric depth recovered from the calibrated stereo rig |
| Streams | Ten: 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, instance segmentation, action labels |
| Sync and IMU | Hardware-triggered onto one PTP clock domain, aligned under 2 ms; 6-DoF IMU at 200 Hz |
| Hand pose | 21 joints per hand, 3D, with a per-joint visibility flag |
| Calibration | Twice per session; accepted only under the 0.28 px RMS reprojection gate |
| Capture location | Firsthand contributor settings, not a customer’s own facility |
| Which tier applies to this row | Not stated; confirmed in the skill spec |
Every collection runs through the same seven-stage end-to-end custom collection pipeline.
Task rows
| Catalog row | Environment | What the catalog states | Skill | Capture mode | Status |
|---|---|---|---|---|---|
| Stir + flip on stovetop | Working Kitchen Set (food prep) | Stir a pan and flip contents. | Medium | Multiview rig + handheld grippers | Available on request |
| Crack + whisk eggs (nearby, context only) | Working Kitchen Set (food prep) | Crack eggs and whisk, then pour to pan. | Medium | Multiview rig + handheld grippers | Available on request |
| Measure + mix (nearby, context only) | Working Kitchen Set (food prep) | Measure ingredients and mix a batter. | Medium | Multiview rig + handheld grippers | Available on request |
| Knead + roll dough (nearby, context only) | Working Kitchen Set (food prep) | Knead dough and roll it flat. | Medium | Multiview rig + handheld grippers | Available on request |
| Assemble a salad (nearby, context only) | Working Kitchen Set (food prep) | Tear greens, add toppings, dress, and toss a salad. | Medium | Multiview rig + handheld grippers | Available on request |
| Peel + chop vegetables (nearby, context only) | Working Kitchen Set (food prep) | Peel and chop vegetables with a knife. | Medium | Multiview rig + handheld grippers | Available on request |
| Recipe, contents, pan or utensil type, heat setting, contents-state or flip-success labels | None | Not stated as rows | Not stated | Not stated | Not stated |
Page comparison
| Page | Main topic | Catalog rows it covers | Status as listed | Capture mode as listed |
|---|---|---|---|---|
| This page: egocentric stovetop stirring and flipping data | Stirring a pan and flipping its contents at a stovetop | Stir + flip on stovetop | Available on request | Multiview rig + handheld grippers |
| /custom-collection/egocentric-cooking-video-data | First-person food-prep and cooking capture in kitchens; stovetop transfer is one category | The kitchen task categories the site documents; not a single row | Per row; see the catalog | Head-mounted capture with hardware-synced streams, as that page states |
| /custom-collection/egocentric-egg-cracking-and-whisking-data | Cracking and whisking eggs, then pouring to a pan | Crack + whisk eggs | Available on request | Multiview rig + handheld grippers |
| /custom-collection/egocentric-baking-and-dough-handling-data | Kneading, rolling dough and mixing batter | Knead + roll dough; Measure + mix | Available on request | Multiview rig + handheld grippers |
| /custom-collection/stereo-cutting-and-food-prep-data | Peeling and chopping vegetables with a knife, framed as stereo capture delivered as MCAP on that page | Peel + chop vegetables | Available on request | Multiview rig + handheld grippers, as the catalog lists it |
Capture challenges
| Challenge | What goes wrong | Stated stream or label |
|---|---|---|
| Stir and flip share one row | Two actions sit in one description | Action segments with verb and noun split each step; the skill spec names each; no stir pattern or flip count is stated |
| Fingers behind objects | Hand occlusion in manipulation | 3D hand pose, 21 joints per hand, with a per-joint visibility flag |
| Wet or reflective surfaces, steam | Depth degrades | Per-frame confidence maps mask low-confidence regions; how the contents of a pan behave in depth is not stated |
| Contents that change shape and state | Masks propagated frame to frame do not track them well | Per-frame instance masks, stored at 848×480 |
| The manipulated object leaves the frame | The episode loses its subject | Reject if the object is out of frame for more than 15% of an episode; the spec names the manipulated object |
| Judging a finished flip | A finished result still needs a label | Not stated: no contents-state, stir-progress or flip-success labels are listed |
Delivered vs not stated
| Item | Status | Detail |
|---|---|---|
| Action segments | Delivered | start_ns, end_ns, verb, noun, one per step |
| 3D hand pose | Delivered | 21 joints per hand; per-joint visibility flag for occluded fingers |
| Contact events | Delivered | The moment a hand meets an item |
| Per-frame instance masks | Delivered | Stored at 848×480 |
| Consent and anonymization records | Delivered | Consent artefact IDs per episode; anonymization record shipped with the data |
| Recipe, contents, pan and utensil types, heat settings | Not stated | Agreed in the skill spec |
| Contents-state, stir-progress, flip-success or success labels | Not stated | Not among the listed labels |
| Cycle times, error rates, and counts of sites, episodes and hours | Not stated | Set per program in the skill spec |
| Force, tactile or thermal readings; safety and certifications | Not stated | Not among the listed streams or claims |
Example footage
One real clip already published on this site, shown as a format and rig example: a head-mounted stereo capture with hands in frame. It does not show stirring, flipping or a stovetop, and it is not an example of the multiview rig with handheld grippers this row uses.
No stirring or flipping clip is published on the site, so this is the nearest clip. Request a sample to see delivered files. See stereo sample page, stereo + wrist-cam example videos, request an egocentric data sample.
FAQ
Sources
Tell us the steps you need, their order, where an attempt starts and ends, what counts as complete and what the camera must see. We will quote reach, timeline and price against a written spec, with a first validated batch in a median of 48 hours where coverage is deep.