Short answer
Routine building tasks such as replacing HVAC filters, lamps, faucet parts and door hardware are good candidates for robot data because they recur and have clear outcomes. Property operators hold the work orders and can allow recordings in their buildings. Work orders give frequency and parts; video gives visual process; neither supplies robot actions, so hands-on robot data is still required.
Why building maintenance is a practical target
A property operator runs thousands of small, similar jobs: swap a filter, replace a lamp, tighten a hinge, clear a drain, change a ceiling tile. Each is short, bounded by a work order and carried out in spaces that already have lighting and power. Compared with industrial repair, the hazards are lower and the parts are standardized. For a team deciding where a humanoid or mobile manipulator might first earn its place, facility tasks sit near the top of the list.
They are also deceptive. A filter swap involves ladders or access panels, orientation arrows, size matching, disposal and a record of the change. Reaching a ceiling diffuser means working above shoulder height with limited vision. This page lays out the tasks and the data each needs. It does not replace the specialized pages on HVAC service data, plumbing repair and installation or electrical installation tasks.
See also the maintenance hub and the humanoid data sourcing guide.
Task anatomy
| Task | Main steps | Typical failure modes |
|---|---|---|
| HVAC filter replacement | Locate unit, open access panel, note airflow arrow and size, remove old filter, seat new filter, close panel, log | Wrong size, reversed arrow, gap at frame, dropped dust, panel latch jammed |
| Lamp or fixture replacement | De-energize, remove cover, extract lamp, check socket, insert new lamp, restore | Wrong base type, broken lamp, loose cover, switch left on |
| Faucet or aerator service | Shut valve, unscrew aerator, clean or replace, reassemble, test | Cross-threading, seal missing, leak after reassembly |
| Door hardware adjustment | Identify closer or hinge, loosen, adjust, tighten, test cycle | Stripped screw, wrong adjustment direction |
| Ceiling tile and grid access | Lift tile, inspect plenum, replace damaged tile | Tile breakage, debris fall, ladder instability |
| Drain and trap clearing | Place bucket, loosen trap, clear, reassemble, flush | Misaligned seal, spill |
Objects are mostly standard: filters in listed sizes, lamps by base type, fasteners by head type. That is what makes the tasks tractable. Environment varies: equipment rooms, rooftops, ceiling voids, tenant spaces.
What each data type contributes
Work orders. A work order records the request, location, asset, time, craft, parts and completion note. Collections of them show which tasks are frequent, how long they take and which parts recur. MaintNet, which includes facilities-domain logbook data alongside aviation and automotive, also ships abbreviation lists and pre-processing tools because maintenance text needs normalization before use (Akhbardeh et al., 2020). Buyers can use work order histories to prioritize which tasks to record.
Preventive maintenance schedules. They show intervals and checklists for filters and fixtures; see preventive maintenance checklists.
Video of technicians. First-person or third-person recordings of the tasks above, ideally with step labels. Existing recordings are rare in routine facility work, because the work is not usually filmed. Training videos by property managers or vendors may exist.
Images of fixtures and assets. Photos of installed filters, labels and fixtures help perception, but they do not cover handling.
What the data can support, and what it cannot
Work orders can support task selection, vocabulary, retrieval and evaluation questions. Video can support visual representation, step recognition and progress estimation. Combined, they can anchor a task library.
Neither contains robot joint trajectories, grasp forces or the contact behavior of a latch or a frame seated in a rail. A humanoid performing a filter swap needs those signals from robot demonstrations, teleoperation or instrumented capture. Human demonstration video does not carry actions, and turning it into robot control uses retargeting, inferred actions or co-training, discussed in the human demonstration video page.
Also absent: the layout knowledge of a particular building. A model trained on one operator's equipment rooms may not recognize another's, so cross-site coverage is a data requirement rather than a nicety.
Sourcing routes
Facility managers and property operators are the natural suppliers. They can provide work-order exports, PM templates and, with permission, access for new recordings. Contract facility-services firms hold similar records for many buildings, which helps with diversity but complicates rights, because the work orders belong to different building owners.
New recording in buildings requires tenant notice, scheduling around occupancy and attention to areas where cameras are inappropriate: restrooms, residential units, medical spaces. The recording in hazardous spaces page covers rooftops and ceiling voids. For lubrication and cleaning variants of these jobs, see cleaning and lubrication tasks.
Illustrative capture and record package
# Illustrative only - synthetic values
task: hvac_filter_replacement
building_type: office_midrise (site_031)
access: ceiling_return_grille | rooftop_unit | air_handler_room
video:
egocentric: 1 chest-mounted, 1080p
third_person: 1 fixed, covers ladder and reach
labels:
steps: [locate, open_panel, read_size_and_arrow, remove, seat, close, log]
filter_size_label: legible_in_frame (yes/no)
outcome: correct_orientation | reversed | gap | wrong_size
height_above_floor_band: floor | reach | overhead
work_order_link: wo_88123 (request text, asset id, parts, minutes)
metadata: operator_A12 (pseudonymous), tenant_notice_given, privacy_zones_masked
Seasonality, access and building variation
Building tasks cluster in time. Filter changes follow seasonal schedules, lamp failures spread evenly, and drain problems peak in specific weather. If your plan depends on task frequency, ask for work orders across at least a full annual cycle so seasonal tasks appear. Access conditions also change within a single task title: the same "replace filter" request might mean a wall grille at head height at one site and a rooftop unit behind a locked hatch at another.
Record the dimensions that will matter to a robot: clearance around the unit, height of the working area, whether the access panel is hinged or screwed, and whether a ladder is required. These are rarely in work orders, so plan to add them during annotation or capture, and compare building types so one portfolio's habits do not dominate.
Quality and acceptance criteria
- Work orders linked to recordings by identifier, with times that agree.
- Filter size and orientation arrows legible in the close view.
- Variation across at least the building types and access types your plan names.
- Recovery events (stuck panel, wrong size) present at a stated share, not only clean runs.
- Evaluation split held out by building, not by clip.
- Tenants, visitors and unrelated workers excluded from frame or handled according to the agreement.
Limitations to plan for
Overhead work puts the camera at an angle that few humanoid heads share. Ladder use adds safety constraints that a robot would handle differently. Buildings differ in control types, access panels and fixture brands, so a dataset from one portfolio may not transfer. Free-text completion notes such as "done" or "replaced" say little about difficulty. And a work order says a task occurred, not how well.
Questions to ask suppliers
- How many buildings, building types and years do the work orders cover?
- Can work orders be linked to specific recordings or photos?
- What access types (ladder, rooftop, plenum) appear in video?
- Were tenants or occupants notified, and were private areas excluded?
- Do work orders include parts, labor time and callbacks?
- What portion of closure notes are free text, and in what style?
- Who owns the work-order data: the property owner, the service firm or both?
- Can new recordings be arranged in additional buildings, and under what conditions?
Discuss your requirements
If you need routine building maintenance data, describe the tasks, building types and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses, and feasibility of any new recording is discussed case by case. A request does not guarantee a match. Related catalog entries cover maintenance logs and field service maintenance records.
Describe the workplace demonstration data your robotics project needs
Sources
- MaintNet: A Collaborative Open-Source Library for Predictive Maintenance Language Resources — facilities-domain logbook data and normalization tools.