Egocentric Server Hot-Swap and Component Seating Data

Egocentric server hot-swap and component seating data is video of the moment a part goes into, or comes out of, a server: a caddy in a bay, a module in a chassis, memory in a slot, a card in its socket. It is recorded against a written spec, with depth, contact events and frame-accurate action segments aligned on one hardware clock where the spec includes them. The tasks are the ones the site states in its task catalog, five of the twelve rows in the Data Center & IT Hardware Bench environment.

The broad data center page covers all twelve rows, including cabling, rails and labels. This page narrows to the rows whose last step is a seat, a latch or a release, because those steps share failure modes that a policy trained on them has to see.

Five rows, all in the Data Center & IT Hardware Bench environment under the Office & IT sector, which the catalog describes as a server rack plus a hardware bench with modules, cabling and drivers: "Hot-swap drive caddy", "Swap PSU module", "Swap fan module", "Seat RAM DIMMs" and "Reseat PCIe card". Four are rated Easy and one, "Reseat PCIe card", is rated Medium.

All five are listed as Available on request and list the same capture setup, a multiview rig with handheld grippers. The site does not state that any of them has been collected and delivered, and this page claims no delivered hours.

Three rows swap a module. The catalog descriptions read "Remove and insert a drive caddy in a bay", "Replace a hot-swap power supply" and "Release and replace a hot-swap fan". All three are rated Easy.

The names tell you where the video should look: the bay, the chassis opening and the release point. Server models, bay counts, module types and the order of the steps are Not stated, so a spec that depends on any of them has to name them.

Two rows seat a part in a socket. "Seat RAM DIMMs" reads "Insert memory modules into slots until they click and latch" and is rated Easy. "Reseat PCIe card" reads "Unbracket, reseat, and secure an add-in card" and is the one Medium row of the five.

The catalog gives the end of the memory step a physical description, a click and a latch, but it states no way of labelling or measuring a seated result. Slot counts, module counts and card types are Not stated.

The egocentric data center and IT hardware page is the broad page: all twelve rows, from cable dressing to asset tags. The stereo data center page covers the stereo and container side of the same twelve rows. This page is narrower than both, five rows and the seating and swapping failure modes only, and it is not a stereo page.

The electronics and device repair page covers consumer devices, and the thermal paste page covers one precision-assembly row. Neither covers server hardware, which the table below sets out.

Seven other rows in the same environment are not covered here: "Apply + scan asset tag", "Cable-comb dressing", "Clip cage nuts + panel", "Connect C13/C14 power", "Insert SFP+ + fiber", "Patch RJ45 cabling" and "Screw-mount rack rails". They appear on the broad page and are only context here.

Tasks people often expect in this topic, such as racking a full server, configuring a switch or decommissioning hardware, are not stated as catalog rows.

The catalog describes steps that end in a physical seat: a caddy removed and inserted in a bay, memory pushed home until it latches, a card unbracketed, reseated and secured. The moment that matters is the one where the hand covers the slot. Lighting, rack density, reach and access are Not stated.

What the site does state is what egocentric video puts into the training distribution: hands entering the frame from below, occlusion during grasp and motion blur during reach. RGB alone leaves a model to infer 3D hand structure it cannot see reliably through occlusion and motion blur, which is why the site delivers hand pose with a visibility flag per joint.

The catalog describes the 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 a head camera at 3840 x 2160 and 60 fps for its instrumented rig, but no camera count, resolution or frame rate for these rows, so the capture tier is agreed in the spec.

The first-person head-mounted mode is listed for other catalog rows. Whether it is offered for these five is Not stated, and is settled in the written spec, as is the gripper model and jaw type, which the site does not state.

The five rows do not list stereo. The site offers a stereo head rig as part of its general stereo offer, and the stereo data center page states that a stereo pair is a capture choice agreed in the skill spec and confirmed on a Spec Pilot, not something already running on these rows.

If a policy needs a calibrated pair on a drive bay or a DIMM slot, start from the stereo data center page. This page does not describe these rows as stereo capture.

Each episode follows the Firsthand episode schema: ten streams synced under 2 ms, with head and wrist RGB, depth, 3D hand pose with per-joint visibility, contact and grasp events, and frame-accurate action segments. Hand pose is delivered as handJoints[2][21][3] in metric units in the head-camera frame, aligned to depth and RGB. For seating work, contact events mark when a part meets a bay or slot, and action segments mark each step.

The site does not state which of those streams the staged gripper mode populates for each row. Object labels use the terms your spec sets. Force or tactile data, and success or quality labels for a seated card or a latched module, are Not stated.

The site offers a tactile glove tier as a pilot-ready option. The glove signal is time-synced to a head-mounted stereo camera by hardware timestamps and an IMU, and each glove is calibrated every session.

The site does not state that the glove tier has been used on any of these five rows, and each lists handheld grippers as its capture mode. Treat force and tactile data as Not stated unless a spec adds the glove tier and agrees the fit with you.

The catalog places the rows in a server rack plus hardware bench environment. The site names no data center, colocation facility, customer or operator, and it does not state capture at a customer own site. Environment, access and schedule are agreed in the spec, and the number of sites is Not stated.

Every episode links to a signed contributor release through a documented consent chain. Faces, plates and screens are detected and irreversibly blurred at ingest, a reviewer confirms the result before anything ships, and the anonymization record is delivered with the data. An undocumented bystander is a stated reject criterion. How asset tags, serial numbers and drive or module labels are handled is Not stated, so name them in the spec.

Episodes are machine-checked at ingest against the stated reject criteria, listed in the table below, and a human reviewer then checks the result. The first of them matters most here: the manipulated object may not be out of frame for more than 15 percent of the episode.

The site states no rule for how long a slot or bay may stay hidden behind a hand, and no blur or sharpness rule. A rejected item goes to the reject log and is recaptured at our cost.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The format table below sets each against what its official documentation states.

Every batch is validated against signed accept criteria before it ships, with a reject log, and delivery is rolling, per validated batch. MCAP delivery is described on the stereo data center page.

Where coverage is already deep, the first validated batch lands in a median of 48 hours from signed spec. Thin environments typically take three to four weeks, because vetting operators there is the bottleneck, and we quote that up front.

On-demand top-ups run against the same spec and accept criteria after each training run. The site states no capture timeline for these five rows beyond a quote against your spec.

You describe the swap and seating tasks, the sequence and the capture mode your policy needs. We turn that into a written spec covering tasks, environments, capture tier, annotation depth and accept criteria, then quote reach, timeline and price before any capture begins.

A Spec Pilot, a small and fast first batch, can prove the spec and accept criteria before you scale. Billing is per validated hour, and every collection ships under a perpetual, buyer-owned commercial license.