Depth on small, close parts
Fastening, seating and fitting happen inches from the lens, where a depth error is a missed fit. Stereo gives depth aligned with the RGB frame.
Hardware and light-assembly plants
Egocentric stereo capture for hardware and light-assembly plants.
Hinges, connectors, lenses and panels are small, shiny and repetitive, which is where depth gets lost. We collect custom egocentric stereo capture for door and window hardware, electronics assembly, lighting products and furniture plants, on demand and to a written spec, with calibrated stereo, synchronized streams and a buyer-owned license.
- Environments
- Four hardware and light-assembly plant types
- Capture
- Head-mounted stereo, calibrated per session
- Sync
- Hardware-triggered, one PTP clock domain
- License
- Perpetual, buyer-owned
Requirements
What does hardware and light-assembly work need from a capture?
Repetition without sameness
The same step recurs with small variation. A written spec lists the variation your pipeline is missing, so it is captured on purpose.
Hand pose and contact in 3D
Hand joints, object poses and contact events are annotated against the stereo geometry, not guessed from a single frame.
Sibling pages for the exact task
Electronics, plastic and light assembly, screwdriving and fastening, and furniture assembly each have a task-level page you can read alongside this one.
Definition
What is egocentric stereo capture for hardware and light-assembly plants?
It is video recorded from the point of view of the person doing the work, through two calibrated cameras spaced like human eyes, in plants that build hardware, electronics, lighting and furniture. The camera sees what the worker sees, including hands and parts, and every frame pair carries recoverable depth.
Environments
Which hardware and light-assembly environments are covered?
Door and window hardware plants: typical work includes fitting hinges, handles and locks, fastening and clipping. Electronics assembly plants: seating connectors, routing cable and fastening housings. Lighting products plants: assembling housings, wiring drivers and fitting lenses. Furniture plants: fitting panels, drilling and fastening, and finishing.
These are examples of the work typical of each setting. Your task list comes from your skill spec, and whether recording for a given plant is already running is Not stated on this page.
Why it is hard
Why is light assembly hard for robot learning?
Metal hardware and lens surfaces throw glare, plain panels give a single camera little texture, and small fasteners vanish behind fingers. Every step is short, so a policy has to get the geometry right quickly and repeatedly.
A calibrated stereo pair measures depth rather than inferring it, and wrist views plus 3D hand pose cover the moments the head camera loses sight of the grasp.
Episodes
What does each episode contain?
Streams are a global-shutter 4K/60 head camera with a 120° lens, two wrist 1080p/60 cameras, depth at 848×480 and 30 fps, and 200 Hz 6-DoF IMUs, all hardware-triggered onto one PTP clock domain. Calibration is re-solved with a checkerboard target at the start and end of every session and ships with the episode.
Which streams a given plant program carries is Not stated and is set in the skill spec.
Delivery
How is the data labeled and delivered?
Labels follow the stereo geometry. Episodes ship as RLDS, LeRobot, WebDataset, HDF5 or zarr, with per-stream nanosecond timestamps in the PTP domain. Collections are billed per validated hour, with a visible reject log, and rejected items are recaptured.
Speed
How fast can a custom hardware or light-assembly dataset be collected?
Where coverage is deep, a first validated batch lands in a median of 48 hours from a signed spec. Where it is thin, plan on three to four weeks. Which applies here is quoted against your spec, because the site does not state coverage for these plant types.
Getting started
How do you start a hardware or light-assembly pilot?
Name the plant type, list the steps in order, say how each starts and ends, and describe the variation your pipeline is missing. We turn that into a skill spec, quote reach, timeline and price, and usually begin with a small Spec Pilot. Then request a custom dataset.
Environments
What work is typical of each hardware and light-assembly plant?
| Environment | Typical work (examples) |
|---|---|
| Door and window hardware plant | Fitting hinges, handles and locks; fastening and clipping |
| Electronics assembly plant | Seating connectors; routing cable; fastening housings |
| Lighting products plant | Assembling housings; wiring drivers; fitting lenses |
| Furniture plant | Fitting panels; drilling and fastening; finishing |
Every collection runs through the same seven-stage end-to-end custom collection pipeline.
Delivered vs not stated
What is stated, and what is Not stated?
| Item | Status | Detail |
|---|---|---|
| Stereo calibrated per session | Stated | Shipped with each episode |
| Hardware sync | Stated | One PTP clock domain |
| Hand pose | Stated | 21 joints per hand, 3D, in the head-camera frame |
| Delivery formats | Stated | RLDS, LeRobot, WebDataset, HDF5 or zarr |
| License | Stated | Perpetual, buyer-owned |
| Plant types recording now | Not stated | Confirmed before a quote |
| Part types, tolerances and standards | Not stated | Not promised on this page |
| Cycle times and episode counts | Not stated | Agreed in the spec |
FAQ
Egocentric stereo capture for hardware and light-assembly plants, answered.
- Which plants are covered?
- Door and window hardware, electronics assembly, lighting products and furniture plants. Access for each is confirmed before a quote.
- How does this relate to your task-level assembly pages?
- Those pages cover the task, such as electronics assembly or screwdriving. This page covers capturing it in a plant environment with stereo. Read them together.
- What does an episode contain?
- Head-mounted stereo, two wrist views, depth and IMU on one hardware clock, with calibration shipped alongside and 3D hand pose.
- Are tolerances, standards or cycle times included?
- Not stated. Those are agreed per order and are not promised on this page.
- How fast is a first batch?
- Where coverage is deep, a median of 48 hours from a signed spec. Where it is thin, three to four weeks.
- Who owns the data?
- You do. Licenses are perpetual and buyer-owned.
Scope a hardware or light-assembly capture.
Tell us the plant type, the steps and the failure cases. We quote reach, timeline and price against a written spec before capture begins.