Short answer
Janitorial workflow data records how a cleaning round is organized: the order of rooms and tasks, which supplies are carried, when consumables are replaced and how a worker adapts to what they find. It supports task sequencing and long-horizon understanding. It does not supply the wiping primitive, robot actions or contact forces, so buyers pair it with lower-level manipulation data.
What "workflow-level" means
A janitorial round is a chain of short tasks: enter the restroom, check stalls, empty bins, clean fixtures, wipe mirrors, restock paper and soap, mop out, stage the cart and move on. The surface-contact skill inside each step is covered in wiping, scrubbing and surface cleaning demonstrations. This page is about the layer above it: sequence, state tracking, tool and consumable logistics, and decisions about what to do next.
That layer matters for a humanoid because individual skills are only useful if something chooses and orders them. Workflow data feeds task planning, long-horizon evaluation and language grounding, and it can show which steps are conditional ("restock only if the dispenser is below half").
Task anatomy of a janitorial round
| Stage | Typical content | What a recording should capture |
|---|---|---|
| Route planning | Which rooms, in what order, with what priority | Floor plan or room list, start and end time |
| Cart setup | Chemicals, cloths, bags, paper goods, signage | Cart contents before and after the round |
| Per-room sequence | High-to-low cleaning, dry before wet, clean before dirty | Step boundaries and ordering |
| Restocking | Replace towels, tissue, soap, liners | Item, location, level before and after |
| Waste handling | Bin liner swaps, bag tie-off, transfer to bulk bin | Bag state, weight class |
| Safety signage | Wet-floor signs, door holds | Placement and removal events |
| Close-out | Checklist, issue reporting, cart return | Log entry or sign-off |
Failure modes to look for: skipped steps when the worker is rushed, a restock done before the surface is wiped, a missing consumable that forces a return trip, rooms occupied by guests, and tasks interrupted by a request from a manager.
What the data can support
- Task sequencing and decomposition. Workflow recordings with step labels teach models how a high-level instruction such as "turn the restroom" splits into ordered subtasks. Task graphs derived from procedures make that structure explicit.
- Long-horizon understanding. Rounds last tens of minutes, which exposes a model to state that persists across many steps. See long-horizon task demonstrations.
- Mobile-manipulation context. The traverse between rooms, door handling and cart movement are part of the data. See mobile manipulation demonstrations.
- Inventory and state tracking. Consumable levels before and after a restock create labeled examples of "needs restocking" visual states.
- Evaluation of plan quality. A held-out set of rounds from different buildings can test whether a planner proposes orderings that a supervisor would accept.
What it cannot support
- Low-level control. Video of a worker pushing a cart gives no joint trajectories and no wheel commands. Humanoid or mobile-manipulator control still depends on robot data or teleoperation.
- Forces. Pushing a loaded cart, wringing a mop and pulling a stuck bin liner involve loads that video does not record.
- Unwritten judgment. Why a worker skipped a stall or flagged a leak is only visible if narrated or logged.
- Representative variety. A single contractor's route in one building teaches that building. Layouts, fixtures and dispensers vary widely.
- Proof the procedure matches practice. Written cleaning schedules and what workers do often differ. Procedural drift is a documented diligence concern for any SOP-linked dataset.
Mapping tasks to a taxonomy
If you need a shared vocabulary for steps, the O*NET database includes task statements and detailed work activities, with a file that maps tasks to detailed work activities (O*NET Resource Center). It is a reasonable starting point for naming janitorial activities consistently across suppliers, though its statements are occupation-level and far coarser than robot-ready step labels. Treat it as a naming scaffold, not a segmentation scheme. For a narrate-then-segment annotation approach, EPIC-KITCHENS-100 used a pause-and-talk narration stage followed by transcription, parsing and temporal segmentation (Damen et al., 2020).
Illustrative collection plan for a workflow dataset
# Illustrative only; synthetic values
dataset: janitorial_rounds_v0
scope: commercial restrooms and common areas, 4 buildings, 3 contractors
unit: one complete round (target 15-45 min)
capture:
head_camera: egocentric, constant frame rate, microphone off unless consented
cart_camera: optional fixed view of cart contents at start and end
logs: route sheet, restock log, issue log (as exported by the contractor)
labels:
level_1: room-level segments with room_type
level_2: task segments (empty_bins, clean_fixtures, restock, mop, stage_cart)
level_3: optional primitive tags linking to wipe/scrub clips
state: consumable_level before/after (full, half, low, empty)
deviations: skipped, reordered, interrupted, repeated
splits: held out by building and by operator
rights: license for model training and internal evaluation; deletion on request
Change scope and levels to your architecture. A planner needs levels 1 and 2 with consistent labels; a perception team may care more about the consumable state.
Acceptance criteria
- Round start and end are marked, with no unexplained gaps in the video.
- Task segment labels follow a published vocabulary, with a measured agreement rate between annotators.
- Every restock event carries an item type and a before/after level.
- The route sheet or log can be linked to recordings by ID.
- Deviations are labeled, not removed; a dataset of perfect rounds will mislead a planner.
- Training and evaluation rounds are separated by building and operator.
Suppliers and diligence
Typical suppliers are cleaning contractors and in-house facility managers; see facility-services contractors as demonstration suppliers. They may also hold the written side: cleaning schedules and checklists that fit housekeeping and cleaning SOPs, which can be linked to video.
Recordings of employees in a building may involve other people's personal data too: occupants, guests and visitors. The EDPB's guidelines on video devices are one reference for the kinds of questions regulators ask about recording (EDPB, 2020). This is not legal advice. Contractors also often work in buildings they do not own, so the building owner's permission may matter as much as the worker's. Redaction does not settle that question.
Questions to ask suppliers
- Whose premises were recorded, and did the building owner agree to recording for AI training?
- Are recordings of full rounds, or only selected tasks?
- Is there a route sheet, checklist or log you can link to each recording?
- How many distinct buildings and layouts are covered?
- Are interruptions, skipped steps and rework preserved?
- Who supplied the cleaning procedure, and can that content be licensed with the video?
- Were occupants or guests in view, and how was that handled?
- Could a new recording be scripted around your cart and consumable setup?
Discuss your requirements
If you need workflow-level cleaning data, describe the task sequences, modalities and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX follows up about potential sources.
See the humanoid training data guide for the full map and the workplace task demonstration hub for sibling tasks. The catalog entry on first-person video of skilled manual work and the SOPs and playbooks category describe adjacent material.
Describe the workplace demonstration data your robotics project needs
Sources
- ONET Database (ONET 29.1 per search snapshot): Task Statements and Detailed Work Activities — O*NET task statements and detailed work activities as a taxonomy for janitorial tasks
- Guidelines 3/2019 on processing of personal data through video devices (Version 2.0) — Guidance on personal data processed through video devices
- Rescaling Egocentric Vision — Narration-then-segmentation annotation approach for egocentric kitchen video