Egocentric stereo capture in hard environments.

A hard environment is a workplace where ordinary first-person video loses information a robot needs. Think glossy or dark surfaces, plain panels with little texture, mixed light, tight clutter, and small parts held close to the lens. Plant floors that build appliances, hardware, electronics, lighting and furniture are full of all four.

These are the working conditions robot manipulation has to survive, and they are exactly where a lab most often lacks data. The environments below are examples of where we scope custom stereo collection. Which of them is recorded for your order is agreed in the spec.

We scope eight plant types as example targets: air conditioner plants, gas appliance plants, refrigerator and cold appliance plants, water meter plants, door and window hardware plants, electronics assembly plants, lighting products plants and furniture plants. Each has its own detail page, linked below, with the work typical of that setting.

Whether recording for a particular plant type is already running is Not stated on this page. We confirm access, consent and timeline before we quote, and we say so plainly if a plant type needs a new venue.

The head-mounted camera is a global-shutter 4K/60 unit with a 120° lens. Two wrist-mounted 1080p/60 cameras sit alongside a depth sensor at 848×480 and 30 fps and 200 Hz 6-DoF IMUs. All streams are hardware-triggered onto a single PTP clock domain, and intrinsics and extrinsics are re-solved with a checkerboard target at the start and end of every session so drift is detectable rather than assumed.

Stereo pairs are delivered raw with calibration, or rectified with disparity if you want it added. Hand pose is 3D, in metres in the head-camera frame.

Depth from a single frame is a guess, and a wrong guess about where a part sits costs a failed insertion. In a glare-heavy or low-texture scene the guess gets worse. A calibrated stereo pair removes the guess by measuring disparity between two views taken at a fixed, known baseline.

Stereo also keeps working outdoors where active depth sensors struggle, and many programs capture both and sync them on one clock. The head camera still loses sight of a grasp when the hand blocks it, which is why wrist views and hand pose travel with the stereo pair.

Labels follow the stereo geometry: 3D hand joints, object poses and contact events are annotated against it rather than guessed from one frame. Episodes ship as RLDS, LeRobot, WebDataset, HDF5 or zarr, with per-stream nanosecond timestamps in the PTP domain so sync can be verified independently.

Which labels a given plant program carries is Not stated here. It is set in the skill spec and quoted before capture begins.

Where our coverage is already deep, a first validated batch lands in a median of 48 hours from a signed spec. Where coverage is thin, plan on three to four weeks. Which applies to a given plant type is quoted against the spec, since the site does not state that coverage for these environments.

Collections are billed per validated hour, and a visible reject log shows what did not make the cut and why. Rejected items are recaptured rather than shipped with a caveat.

Describe the plant type, the tasks and their order, how each step starts and ends, and the failure cases your model is missing. That becomes a written skill spec, quoted for reach, timeline and price before any capture. Most programs begin with a Spec Pilot, a small, fast first batch that proves the spec and accept criteria on real footage before volume scales.

Then request a custom dataset and a rep will scope it with you.