Egocentric Window Cleaning Video Data

Egocentric window cleaning data is real-world video recorded from the point of view of a person who is cleaning glass. The camera sits on the head, so the footage shows what the person sees: the hands, the tool and the pane. Alongside the video, each episode carries depth, 3D hand pose, contact events and frame-accurate action segments, all aligned on one hardware clock.

Teams search for it under several names, including window cleaning robot training data, a glass cleaning video dataset and egocentric squeegee task data. They all describe first-person demonstrations of glass cleaning, collected to a written spec. This page covers only what the Firsthand site states about it and marks everything else as Not stated.

The site states two window tasks in its commercial task catalog. The first reads "Wash and squeegee windows using extension poles and squeegees", described as "Wet windows with a pole applicator and pull the water off with a squeegee". The second is "Window & Glass Cleaning", described as "Clean interior and exterior glass with applicators and squeegees and polish edges".

Together those two wordings name six operations: wetting windows with a pole applicator, pulling the water off with a squeegee, cleaning interior glass, cleaning exterior glass, applying with an applicator and polishing edges. Anything outside that wording is Not stated. That includes mirrors as a task of their own, glass doors, display cases, shower screens, vehicle glass and scraping, so none is claimed here. If your policy needs one of them, you write it into the spec and we scope it, because the catalog rows are a starting vocabulary, not a fixed menu.

Both rows are rated Medium. They sit in the Facilities & Cleaning sector, in an environment called Commercial Cleaning & Janitorial, and both list a head-mounted phone or a multiview rig as the capture mode. They differ in status. The pole-and-squeegee row is Active collection, and the Window & Glass Cleaning row is Ready to deploy.

The site defines those statuses plainly. Active collection means recording in the environment is running now, so a spec can join an existing capture stream. Ready to deploy means contributors and site access are set up and capture starts once a spec is signed. Neither status states how many hours, episodes or contributors exist for the row, so a first batch is quoted against your spec.

The Commercial Cleaning & Janitorial environment lists an office building, a shopping mall, an airport and an industrial facility among its examples, with the tasks mopping, vacuuming, surface sanitizing and restroom cleaning. It holds 15 rows. Eight are rated Easy and seven are rated Medium, including both window rows, and no row is rated Hard. Eight rows are Active collection and seven are Ready to deploy.

The site defines two capture modes. A head-mounted phone records the first-person view of real work on site. A multiview rig with handheld grippers records a staged working set. Both window rows list both modes and do not say which applies, so the site does not settle whether a given batch is a real workplace or a staged set. That is agreed in the spec.

The site names no building, venue or company, and does not state the number of sites, the regions, the building types or the heights of any glass. It does not state capture at a customer own facility either. Environment, access and schedule are agreed in the written spec.

A head-mounted phone records the first-person view of the job, and a multiview rig adds synchronized third-person angles when the spec asks for them. On the staged set, the operator performs the task with handheld grippers, so the demonstration maps onto a robot end effector. A wrist-mounted camera is available alongside the head camera for close-range detail on the tool edge. Camera count, resolution, frame rate and pane layout are Not stated.

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. For glass cleaning, hand pose tracks the fingers and wrist as they hold and guide a tool, contact events mark when a hand or tool meets or leaves an object, and action segments mark each pass or step of the job.

Object labels use the terms your spec sets. Window or pane counts, glass types, tool types, cleaning products and building types are Not stated. Task durations, cycle times, throughput and the number of episodes are Not stated too. Success or quality labels, such as whether a pane was left clean, are Not stated, as are certifications and robot results. The site makes no safety-compliance claim for this work, and states no height, ladder or work-at-height setup.

The site rates both window rows Medium, one step above the Easy rows that make up the rest of the environment, and none of its rows is Hard. That rating is the site only statement on difficulty. The site gives no technical analysis of glass cleaning, so this page makes no claim about robot results, and it does not state how depth behaves on glass or reflective surfaces.

What the row wording implies is a sequence of different operations. Wetting a window with a pole applicator, pulling the water off with a squeegee, cleaning interior and exterior glass and polishing edges each use the hands and the tool differently. The viewpoint matters too. The site states that a robot or humanoid perceives from a head- or wrist-mounted camera, not a static overhead shot, and that egocentric capture matches that geometry.

A good spec treats each named operation as its own task. Wetting, pulling water off, cleaning interior glass, cleaning exterior glass and polishing edges each get a start state, an end state and a set of accept criteria. That keeps action segments consistent across contributors and lets you ask for more of one operation without re-collecting the others.

Other pages cover other task families: general wiping, a hotel bathroom row, floor rows, and glass and mosaic craft, which leaves glass polishing out because that is a cleaning task.

This page centres on the two window rows in the Commercial Cleaning & Janitorial environment. The catalog also words glass cleaning inside two larger Hotel & Lodging rows: Bathroom cleaning includes "wipe mirrors and surfaces", and Lobby & Common-Area Cleaning includes "polish glass". Neither is a window row of its own, so this page does not count them as window tasks.

Delivery formats are RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. Window cleaning episodes use the same schema as every other Firsthand collection, so a loader that reads another Firsthand dataset reads this one without changes. The format table below sets each format 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. Rejected items are recaptured at our cost, and billing is per validated hour.

Not in standard capture. Force or tactile readings are Not stated for either window row. The site separately offers a tactile glove tier as a pilot-ready option, in which glove signals are time-synced to a head-mounted stereo camera by hardware timestamps and an IMU. The site does not state that the tier has been used on window cleaning, so agree the fit with us first.

Every episode links to a signed contributor release through a documented consent chain, and the consent record is part of the delivery. A vision model runs on every frame at ingest, flagging faces, heads, licence plates and any legible screen or document. Those regions are irreversibly blurred, and a human reviewer confirms the result before footage ships.

If the reviewer is unsure a region is fully covered, the episode is held or discarded, never downgraded and shipped. The site does not state that signage or text visible through glass is blurred, so raise any such need before the spec is signed. It states no in-home capture for this work.

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. The site does not state how deep coverage is for either window row, so plan on a quote rather than assuming the 48 hour figure.

The site also states program milestones. A pilot reaches its first delivery in two to three weeks, and a standing program reaches full rate in four to six weeks. Those assume capture coverage is already in place. On-demand top-ups run against the same spec and accept criteria after each training run. A batch size and a later-batch turnaround are Not stated, so both are quoted against your spec.

You describe the glass operations, the environment and the sequence 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. This is the custom egocentric data collection route that every Firsthand task family follows.

A Spec Pilot, a small and fast first batch, can prove the spec and accept criteria before you scale. Every collection ships under a perpetual, buyer-owned commercial license, with a reject log and signed accept criteria on every batch.