Egocentric Precision Assembly Bench Video Data

Egocentric precision assembly bench data is video of hands-on bench work with small parts, recorded to a written skill spec so that a robot learning system can see how the work is done. In the task catalog the setting is called Precision Assembly Bench (electronics) and sits in the Manufacturing & Workshops sector. Its example line reads "Assembly bench: PCB/electronics, fasteners, fixtures, solder/crimp tools, small parts". The same idea sits behind any precision assembly robot training data, assembly bench video dataset or egocentric fine assembly data request, applied here as a custom precision assembly data collection scoped to your own steps.

On this site the data is human capture: a paid, consented contributor performs the task while the rig records it. It is not robot or teleoperation data. The catalog lists one capture mode for every bench row, a multiview rig with handheld grippers, which is different from the head-mounted phone mode listed for several other environments. Whether a head-mounted first-person stream accompanies the multiview rig on a bench program is Not stated for these rows and is agreed in the skill spec.

The catalog lists 15 rows for the Precision Assembly Bench (electronics) environment. All 15 are Available on request, which the site defines as scoped and recruited to spec when a buyer asks for it, rather than already set up and recording. Skill ratings are two Easy, ten Medium and three Hard. The three Hard rows are the busbar, limit-block and screw sequence, crimping a terminal, and through-hole soldering.

The first table lists all 15 rows as the catalog states them. The second sorts them into kinds of bench work. Small-part insertion, fastening, fixture loading, connector seating and kitting each have a stated row. Inspection of an assembled part has no row in this environment, so it is Not stated here.

The site states no part numbers, tool models, torque values, tolerances, cleanroom or static-control practices, certifications, standards, cycle times, throughput figures or number of sites for this environment.

Each attempt at a task is delivered as one episode. The site’s episode schema lists action segments, contact events, 3D hand pose for both hands with a per-joint visibility flag, and instance masks, alongside int64 nanosecond timestamps per stream and camera intrinsics and extrinsics. Hand pose is stated as 21 joints per hand, in 3D, in metres in the head-camera frame.

The site’s general capture statement lists a head-mounted global-shutter 4K/60 camera with a 120° lens, two wrist-mounted 1080p/60 cameras, an active-stereo depth sensor at 848×480 and 30 fps, and 200 Hz 6-DoF IMUs, all hardware-triggered onto a single PTP clock domain. It describes ten synchronized streams aligned under 2 ms. Which of those streams a bench program carries is Not stated for this environment and is decided in the skill spec, because the bench rows list a multiview rig with handheld grippers rather than the head-mounted mode.

Force and tactile readings are not among the standard streams. A separate glove tier records a tactile glove signal beside the video and has its own page. Whether it applies to a precision assembly bench program is Not stated and would be agreed in the spec.

Action segments divide an episode into start- and end-timestamped spans, each carrying start_ns, end_ns, verb and noun, aligned to every other stream. For bench work, a segment might be one connector seated, one screw driven or one terminal crimped, as the skill spec defines it. Contact events mark when a hand meets an item; the site does not describe them as force or pressure.

Validation runs against accept criteria written into the spec before capture. A reject log ships with each batch, and only footage that clears the criteria counts as a validated hour. The site states these reject thresholds: a cross-stream sync error above a 2.0 ms ceiling, the manipulated object out of frame for more than 15 percent of the episode, and more than 35 percent invalid depth pixels in the manipulation region. On custom collection engagements, rejected items are recaptured at the provider’s cost, and delivery is billed per validated hour, not per raw hour recorded. These are reject thresholds, not a stated pass rate or yield; no pass rate is stated.

Delivery is in the format your stack already reads: RLDS, LeRobot, WebDataset, HDF5, zarr or Rerun .rrd, with converters shipped as readable source so you can retarget the schema to your own loader. The official RLDS repository describes datasets as episodes made of steps, with metadata and an optional flag for invalid episodes. The official LeRobot repository describes its dataset format as synchronized MP4 videos or images plus Parquet files for state and action data. Whether a bench delivery carries robot action data is Not stated; the site describes human-side capture only.

The license is perpetual, irrevocable and buyer-owned, and non-exclusive by default, with exclusivity available as a separate term. The executed agreement controls.

Bench assembly is short, repeated and tightly constrained. Fingers cover the small part they are placing, and the part covers the fixture it is going into. That is why hand pose carries a per-joint visibility flag, why contact events mark when the hand meets the item, and why instance masks are recorded per frame. The table below maps each difficulty to what the site states.

The catalog rates three bench rows Hard, and it words several rows with "to spec" or "precisely". The site lists no tolerance, torque value, defect label or success label behind those words, so a buyer who needs a particular tolerance or torque has to write it into the skill spec as an accept criterion.

Robot-model research adds two points. The GR00T N1 paper says its human video datasets do not include explicit action labels, and its related-work section says hand-held robot grippers tend to speed up collection but leave a mismatch with the downstream robot. The bench rows pair a multiview rig with handheld grippers so demonstrations map onto a robot end effector; no accuracy for that mapping is stated. This is robot-model research, not Firsthand data, and Firsthand has no relationship with its authors.

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. It does not describe a real workshop for these rows. The site also states that capture happens in Firsthand’s contributor settings, not at a customer’s own facility. Which bench, venue or region a program uses is Not stated and is agreed in the skill spec.

Every contributor signs a written release granting commercial training rights before capture begins, and is paid a published hourly rate rather than a per-task bounty. Consent is tracked through a signed consent chain, and each delivered episode links to it. Identifying details outside the task are anonymized, audio is scrubbed of identifying speech, and a clip where someone cannot be de-identified is discarded rather than shipped. The anonymization record ships with the data.

Safety practice, protective equipment, cleanroom practice, standards and certifications are not stated for this environment, and this page makes no such claim.

From signed spec to first validated batch, the published timeline is a median of 48 hours in environments where coverage is already deep, and three to four weeks where coverage is thin and contributors need to be vetted and recruited first. All 15 bench rows are Available on request, so they are scoped and recruited to spec when asked for. The site does not say whether the bench counts as deep or thin coverage, so which range applies is confirmed when your spec is quoted rather than assumed.

A spec conversation happens the same week, and a tight spec is the quickest route to a fast, on-demand dataset.

Describe the bench steps and their order, the parts, fixtures and jigs involved, what counts as a finished step, and the failure or variation cases your pipeline is missing. That becomes a written skill spec, quoted for reach, timeline and price before any capture begins. Our skill spec guide walks through how to write one.

Most programs begin with a Spec Pilot, a small, fast first batch that proves the spec and the accept criteria on real footage before volume scales. Collections are billed per validated hour.