Short answer
Cleaning and lubrication data usually pairs recorded routine tasks, such as wiping down a machine or greasing a fitting, with schedules that say what is done and when. The schedules give task structure and frequency; the video gives hand-object context and step order. Fluid handling, dosage and contact are not captured as robot actions or forces, so a humanoid still needs robot-collected data.
What makes this task family different
Cleaning and lubrication are the least glamorous maintenance jobs and among the most frequent. They are also where a mobile manipulator might be given routine work first, because the steps are repetitive and the tolerance for variation is wider than in a precision repair. That is a hypothesis for each buyer to test, not an established result.
Three features shape the data. First, the task involves a substance: cleaning agent, grease, oil or water. Fluids change how surfaces look and behave in ways video captures poorly. Second, access matters: fittings are often behind guards, under covers or at awkward heights. Third, the work is scheduled, so a record trail exists for what was supposed to happen, which may or may not match what was recorded on camera.
If your interest is the wider category of recurring checks, the preventive-maintenance checklists page and the operator rounds and walkdowns page cover the document and walk-through sides. This page stays with the hands-on cleaning and lubrication steps. The maintenance cluster hub shows neighboring tasks, and the industrial machinery maintenance page covers the equipment these routines usually serve.
Task anatomy
Ask the supplier's technicians to confirm the real sequence for each equipment type. A working decomposition for discussion:
- Scope the job. Reading a lubrication schedule or route card to find points, intervals and the specified product. The schedule is a record; the lookup is a perception and language problem.
- Prepare the machine. Many cleaning and lubrication jobs require the machine to be stopped and safe to approach. Federal rules in the US describe employers' duty to develop, document and use procedures for controlling hazardous energy when employees service equipment (OSHA, 29 CFR 1910.147). Whether a given task falls under such procedures is a question for the site's safety staff, not for this page.
- Locate and open. Finding fittings, reservoirs, sight glasses or filters; removing covers.
- Clean. Wiping, brushing, degreasing, vacuuming. The visual outcome (is the surface clean?) is a classification target.
- Apply. Grease gun, oil can, spray or reservoir top-up. The amount applied is typically judged by sight, feel or counted strokes, depending on the equipment.
- Verify and record. Checking levels, wiping excess, logging completion.
Failure modes include missed points, wrong product, over-application, contamination from dirty tools, and skipped steps when time is short. A dataset with only clean passes cannot teach a model what a missed point looks like.
What the data can and cannot support
| Source | Can support | Cannot support |
|---|---|---|
| Lubrication schedules and route cards | Task structure, intervals, point naming, product vocabulary | How anything is physically done |
| Completion logs | Completion patterns, missed-task language | Quality of execution |
| Egocentric demonstration video | Step recognition, hand-object understanding, language grounding | Fluid volume, applied pressure, robot motions |
| Fixed-camera video | Context, sequence, duration | Fine detail when hands are hidden |
| Instrumented or teleoperated capture | Action and force signals for control training | Breadth across sites |
A model can learn from video what grease application looks like; it cannot learn how to squeeze a grease gun against a particular fitting without contact data. See what human demonstration video can and cannot do for robot learning and instrumented tools for what sensors add.
Illustrative collection package
This is a synthetic example. Replace every value with your own requirements.
# Illustrative only
task_family: lubrication_round
assets: [asset_0042, asset_0043] # synthetic IDs
points_per_round: buyer_defined
modalities:
video: head_mounted + one fixed context camera
records: [route_card_text, completion_log]
labels:
- point_id_per_segment
- product_name_as_written
- step_boundaries (locate, clean, apply, verify)
- missed_or_repeated_point_flag
- surface_clean_vs_soiled (per frame range)
metadata: [site_id, operator_id_pseudonymous, shift, lighting_condition]
excluded: [robot_actions, applied_force, fluid_volume_measurement]
Add scale references, such as a visible gauge or marked reservoir level, if you want to estimate quantities from video. Without them, treat fluid amounts as unobservable.
Supplier types and sourcing routes
Plants, fleet depots and facilities teams run cleaning and lubrication every day, and their schedules are often already in a maintenance system. Existing video of this work is rare, because it is not usually recorded. That points toward commissioned recording at a cooperating site, with records licensed alongside it. See existing assets, new collection or specialist partners for how to weigh the routes, and the hand tool use demonstrations page for the tool side of the job.
Recording near moving or pressurized equipment raises site-safety conditions that the host site sets. Recording people raises consent questions; see worker consent and identifiable people. This is not legal advice.
Limits and extra preparation
- Fluids are poorly observed. Transparency, glare and residue confuse vision models. Ask how the supplier handled it, if at all.
- Schedules describe intent. A route card says what should happen, not what the technician actually did.
- Equipment variety. Points, fittings and covers differ by machine; one site's route does not generalize automatically.
- Embodiment. Reaching into a machine with a human arm is not the same as doing so with a humanoid arm and hand. See the embodiment gap page.
- Safety logic. Whether and when to work on a running machine is a deployer's rule, not a lesson in footage; see the lockout/tagout procedures page for the document type.
- Evaluation. Hold out sites and machines if you claim generalization.
Questions to ask suppliers
- Which equipment types and lubrication points does each recording or record cover?
- Are schedules linked to completion logs and to any video of the same round?
- Was the work recorded as it happened, or performed again for the camera?
- Where does the camera sit when hands are inside guards or under covers?
- Are missed points, rework or wrong-product events preserved?
- How are products named: brand, grade, or internal shorthand?
- Which energy-control or access steps appear on camera, and which happen off camera?
- Did the site restrict recording near any machine, and how does that affect coverage?
- Who owns the schedules and logs: the site, a maintenance contractor or a software vendor?
Discuss your requirements
If cleaning and lubrication tasks belong in your maintenance data plan, describe the equipment, modality and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources; whether a commissioned collection is feasible is discussed case by case. A request does not guarantee a match. For the full map of supply routes, start with the humanoid training data guide.
Describe the workplace demonstration data your robotics project needs
Sources
- 29 CFR 1910.147 The control of hazardous energy (lockout/tagout) — employers' duty to develop, document and use hazardous-energy procedures when servicing equipment; not specific to cleaning or lubrication.