One row, four verbs, three hand tools
The row names handling, cutting, forming and finishing, with tongs, cutters and grinders. A spec says which of those a collection records and in what order, rather than relying on one label.
Now recruiting capture teams worldwide. Get paid to record first-person video of real work.
Apply to captureHard task
Requirements
The row names handling, cutting, forming and finishing, with tongs, cutters and grinders. A spec says which of those a collection records and in what order, rather than relying on one label.
A stereo pair at human-eye spacing makes depth recoverable from each frame pair, so the glass, the tool and the hands holding them share one frame.
Pieces are small and are often hidden by the fingers that hold them. 3D hand pose carries a per-joint visibility flag, so a finger hidden behind a piece is flagged rather than guessed.
Streams are hardware-triggered onto one clock with a 2.0 ms cross-stream ceiling. MCAP can hold several timestamped channels in one file; which streams go in it is set in the skill spec.
Definition
It is first-person video of a person doing the work the catalog row describes, handling molten or sheet glass and then cutting, forming and finishing pieces with tongs, cutters and grinders, recorded from their own point of view with two calibrated cameras and with hand pose and sensor streams alongside. Stereo describes the capture, and MCAP describes the container. They are separate choices.
The data is human capture by paid, consented contributors, not robot or teleoperation data. It is built for teams training molten and sheet glass handling robot policies, or a glass cutting and finishing dataset for a model, who want depth and hand pose recorded rather than inferred from a single view. Firsthand is not affiliated with Foxglove or the MCAP maintainers.
Catalog row
One. In Heavy Industry & Fabrication, in the Industrial & Manufacturing sector, whose listed examples include metal fabrication, woodworking, glass manufacturing, a concrete plant and a welding workshop, "Glass manufacturing" is described as "Handle molten or sheet glass, cut, form and finish pieces with tongs, cutters and grinders." It is rated Hard, its capture setup is Phone (head-mounted) / multiview rig, and it is listed as Available on request.
The site says nothing further about what that status 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. The first table below reproduces the row exactly as the catalog states it. This page names only the tasks, tools and material that appear in that row.
Hard
Only what the catalog says. Hard is the value in the Skill column for this row, and the site does not publish a separate rubric, failure rate or tolerance for it, so this page does not invent one. The row itself shows what is demanding in plain terms: it chains four verbs, handling, cutting, forming and finishing, and it pairs them with three hand tools, tongs, cutters and grinders, on molten or sheet glass.
The site also notes that this is hand-tool glass work, not machine controls, which is why it sits apart from the Operating Heavy Machinery Controls row in the same environment. The Hard-task hub lists this row beside the other Hard rows and links to the page for each. Nothing here states how skilled any contributor is, how often a step fails or how a finished piece is judged.
Difference
The egocentric glass manufacturing page covers the same one row as egocentric capture and refers to the MCAP hub for the stereo rig. The stereo glass cutting and forming page treats the Glass manufacturing row together with the four Glass & Mosaic Studio rows, and the stereo glass etching and glassblowing page treats two studio rows. This page is the stereo and MCAP treatment of the one Glass manufacturing row on its own.
The site does not list an off-the-shelf glass dataset, so this work is custom collection. The comparison table below sets the four pages side by side.
Why stereo
Glass work is about where the hands, the piece and the tool sit relative to one another: where a cutter meets a sheet, how tongs hold a piece, how a grinder meets an edge. A stereo pair recovers depth from the calibrated frame pair rather than from a standalone depth sensor, so the glass, the tool and the hands share a frame.
The site gives no depth accuracy figure and no label for a finished cut, a formed piece or a finished edge, so whether those are resolvable in depth is Not stated. The site states elsewhere that wet and reflective surfaces degrade active-stereo depth. How transparent or reflective glass behaves is Not stated, and a glass spec should expect the same care and check it on a Spec Pilot.
Capture setup
The live row lists a head-mounted phone or a multiview rig as its capture setup. It does not list stereo, so the site does not claim that stereo footage of glass manufacturing exists. A stereo pair is available through the site’s general stereo offer and is a capture choice agreed in the skill spec and confirmed on a Spec Pilot.
The site describes two stereo setups and does not say which one a glass order uses. The hard-environments stack is a global-shutter 4K/60 head camera with a 120° lens, two 1080p/60 wrist cameras, a depth sensor at 848×480 and 30 fps, and 200 Hz 6-DoF IMUs. The off-the-shelf stereo head rig, dataset 06, is dual fisheye cameras at 1080p per eye, a 62 mm baseline, per-unit calibration, a 9-axis IMU and hand-tracking output.
Intrinsics and extrinsics are re-solved at the start and end of every session, and streams are hardware-triggered onto one clock with a 2.0 ms cross-stream ceiling. The catalog states that a multiview rig adds synchronized third-person angles when the spec asks for them, and wrist cameras give a closer view of the hands. Capture at your own facility is listed as not offered.
MCAP
MCAP is an open source container file format for multimodal log data, supporting multiple channels of timestamped pre-serialized data. It is serialization-agnostic, so messages can be Protobuf, DDS (CDR), ROS, JSON and more. Its design is row-oriented and append-only, it stores schemas alongside the data, and it offers an optional index plus optional LZ4 or Zstandard chunk compression.
Reader and writer libraries exist for C++, Go, Python, Rust, Swift and TypeScript, and the mcap CLI can inspect, merge and split files. The foxglove/mcap repository is MIT-licensed; that is the MCAP project’s licence, not the licence of a Firsthand delivery. The MCAP hub page and the glossary entry cover the format in full.
Streams
Firsthand offers MCAP delivery, and dataset 06 is stated as shipping as MCAP. For glass work the site states hand pose, a stereo-derived depth stream with confidence maps, IMU, contact events, action segments and per-frame instance masks elsewhere. What it does not state is which of those are channels in an MCAP file, how they are named, what encoding they use or at what rate.
No topic name, message schema or rate is stated on this page, and none should be assumed. Name the streams that matter in the skill spec and confirm them on a Spec Pilot batch. The stream table below sets each stream beside what the site states about it.
Failure modes
Hands are the first risk. Pieces are small and are often hidden by the fingers that hold them, and a tool can hide the edge being worked. The visibility flag on each joint is the answer the site gives, contact events mark when a hand meets the item, masks are recorded per frame, and a spec can ask for wrist views so a grip the head camera loses is still recorded.
Surfaces are the second. Glass is transparent or reflective, and the site states that wet and reflective surfaces degrade active-stereo depth and that depth confidence maps mask low-confidence regions. Behaviour on glass itself is Not stated. Thermal, force and tactile readings are not among the listed streams either, and this page claims none.
Calibration is the third, because stereo depth is only as good as the rectification behind it: calibration is measured at the start and end of every session under a 0.28 px RMS gate. An episode is also rejected when the manipulated object is out of frame for more than 15 percent of it or an undocumented bystander appears, and a reject log ships with every batch.
Safety and privacy
Nothing beyond the row. Safety, protective equipment, temperatures, glass types and certification are Not stated for this row, and this page makes no such claim. What a contributor may wear, do or show is agreed in the spec.
For every delivered episode the site states that faces, plates and screens are blurred irreversibly, and that footage that cannot be de-identified is discarded. Whether finished pieces or original designs are blurred or protected beyond the licence is Not stated.
Fast · on demand
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 stereo glass manufacturing order, and the row is Available on request, so recruitment comes first and the timeline is quoted against your spec before capture begins.
The site lists no off-the-shelf glass dataset, so this is custom collection: custom collection at scale through a network of thousands of capture providers, for a consistent, recurring data flow to labs. The only stereo head rig the site states as shipping as MCAP is dataset 06, and this page does not claim off-the-shelf stereo coverage of glass work. On-demand top-ups run against the same spec and accept criteria after each training run.
Ordering
Describe the steps and their order, the glass, tongs, cutters and grinders and their start state, what counts as a finished cut, form or edge, and the variation your pipeline is missing. Then name the streams you want in the file, the topic naming and message encoding your reader expects, whether calibration ships inside the file or beside it, and whether MCAP is the only delivery or sits beside RLDS, LeRobot or WebDataset.
Most programs begin with a Spec Pilot, a small first batch that proves the spec and accept criteria on real footage before you scale. Billing is per validated hour, rejected items are recaptured at the provider’s cost, and every collection ships under a perpetual, buyer-owned licence. The spec table below shows what each element decides.
At a glance
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 row
| Task (as listed) | Description (as listed) | Environment | Skill | Capture setup | Status |
|---|---|---|---|---|---|
| Glass manufacturing | Handle molten or sheet glass, cut, form and finish pieces with tongs, cutters and grinders. | Heavy Industry & Fabrication | Hard | Phone (head-mounted) / multiview rig | Available on request |
Compared pages
| Page | What it covers | Treatment of the Glass manufacturing row |
|---|---|---|
| This page: stereo glass manufacturing data | The one Glass manufacturing row | Stereo pair and MCAP delivery, stream map and failure modes for this row only |
| Egocentric glass manufacturing data | The one Glass manufacturing row | Egocentric treatment; refers to the MCAP hub for the stereo rig |
| Stereo glass cutting and forming data | The Glass manufacturing row and the four Glass & Mosaic Studio rows | Stereo and MCAP treatment of all five rows together |
| Stereo glass etching and glassblowing data | Studio rows in Glass & Mosaic Studio | Not covered; a different environment |
Stated vs not stated
| Item | What the site states | For a glass manufacturing order |
|---|---|---|
| Catalog capture setup | Phone (head-mounted) / multiview rig on the row | A stereo pair is a spec choice; stereo footage of this row is not claimed |
| Stereo head rig (dataset 06) | Dual fisheye, 1080p per eye, 62 mm baseline, per-unit calibration, 9-axis IMU, hand-tracking output; shipped as MCAP | Whether a glass order uses it: Not stated |
| Hard-environments stack | 4K/60 head camera, 120° lens; two 1080p/60 wrist cameras; depth 848×480 at 30 fps; 200 Hz 6-DoF IMU | Whether a glass order uses it: Not stated |
| Reflective surfaces | Wet and reflective surfaces degrade active-stereo depth; depth confidence maps mask low-confidence regions | Behaviour on transparent or reflective glass is Not stated |
| Objects, hands and masks | Per-frame instance masks stored at 848×480; per-joint visibility flag on hand pose; contact events | Stated |
| Calibration | Measured every session and shipped with the episode; 0.28 px RMS reprojection gate | Stated |
| Synchronization | Hardware-triggered onto one clock; 2.0 ms cross-stream ceiling | Stated |
| Row status | Available on request | No volume, hours or delivered stereo footage is claimed on this page |
| Thermal, force or tactile streams | Not among the listed streams | Not claimed on this page |
| Glass types, thickness, piece sizes, defect classes, finished-cut labels, equipment models, safety, protective equipment, temperatures, throughput, accuracy, certification or results | Not stated; the row says molten or sheet glass, tongs, cutters and grinders only | Agreed in the skill spec, or not claimed on this page |
| MCAP topics, schemas, rates | Not stated | Agreed in the skill spec |
Stream map
| Stream or item | What the site states | In an MCAP delivery, as the site states it |
|---|---|---|
| Stereo pair | Dual fisheye, 1080p per eye, 62 mm baseline (dataset 06) | Dataset 06 is stated as shipped as MCAP; channel layout not stated |
| Depth with confidence maps | Recovered from the calibrated stereo pair; confidence map masks low-confidence regions | Not stated |
| Hand pose | 21 joints per hand with a per-joint visibility flag | Channel layout and joint order not stated |
| IMU | 9-axis (dataset 06); 6-DoF at 200 Hz (hard-environments stack) | Channel layout and rate not stated |
| Contact events, action segments, instance masks | Listed in the site’s episode schema; masks stored at 848×480 | Not stated |
| Task instruction | One per episode | Not stated |
| Calibration | Measured every session and shipped with the episode | Placement in the file not stated |
| Topic names, schemas, encoding, rates | Not stated | Agreed in the skill spec |
Skill spec
| Spec element | What it decides | Where it shows up |
|---|---|---|
| Steps and order | What one step is: handle, cut, form, finish | Episode boundaries and action segments; set per program |
| Glass, tools and start state | Which glass, tongs, cutters and grinders, and how they begin | Set per program; the site states no finished-cut label |
| Episode boundary | One cut, one piece or a full sequence | Set per program; the site does not state an episode length for this row |
| Rig and streams | Stereo pair, wrist views, hand pose, IMU and any others | The channels you can read; confirm on the Spec Pilot |
| MCAP topics and encoding | What your reader expects | Set per program |
| Calibration delivery | Inside the file or beside it | Set per program |
| Delivery format | How the data is packaged | MCAP alone, or beside RLDS, LeRobot, WebDataset, HDF5 or zarr |
| Accept criteria | What passes QA | Reject log per batch |
| Privacy | How people and items in frame are protected | Faces, plates and screens blurred at ingest |
| License | Who owns the dataset | Perpetual, irrevocable, buyer-owned, non-exclusive by default |
Formats
| Format | What the official documentation states | Used for here |
|---|---|---|
| MCAP | An open source container for multimodal log data with multiple channels of timestamped, pre-serialized data (mcap.dev) | Stereo delivery in one file; see the MCAP hub page |
| RLDS | A dataset is retrieved as episodes, and each episode holds a dataset of steps (RLDS repository README) | Episode-based training loaders |
| LeRobot | LeRobotDataset v3.0 is a standardized format for robot learning data, including multi-camera video, with many episodes stored per Parquet or MP4 file (LeRobot documentation) | LeRobot dataset format and tooling |
| WebDataset | Tar files in which files that belong together share a basename, split into numbered shards (WebDataset README) | Sharded streaming for large training runs |
| HDF5, zarr and Rerun .rrd | Stated by Firsthand only | Array storage and synchronized playback for review |
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 glass, a cutter, tongs or any glass task.
No glass manufacturing clip is published on the site, so this is labelled as a format and rig example. 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 and their order, the glass, tongs, cutters and grinders and their start states, the streams you want in the file and whether MCAP sits alone or beside a training format. We will quote reach, timeline and price against a written spec.