Preventive Maintenance Checklists

Why preventive maintenance is a natural early target

Scheduled maintenance is the opposite of a one-off repair. The asset is known, the task recurs on a calendar or meter interval, the steps are written down, and the expected result is defined. Diagnosis is minimal. That makes it a plausible first domain for a humanoid or mobile manipulator, and it makes the checklist the cheapest document to obtain.

A checklist alone, however, is a plan, not a performance. It tells a buyer which tasks exist and how often, so a program can rank candidates by frequency, duration and risk. It does not tell a robot how to open a panel, how much force a latch needs, or what a worn belt looks like. For those, see repair video demonstrations and the guidance on prioritizing maintenance tasks for collection.

This page explains how to source checklist data, what fields to demand and where it ends. The maintenance hub and the humanoid data sourcing guide give wider context.

Where checklists come from

Computerized maintenance management systems. A CMMS stores PM templates against asset classes and generates dated work orders. Exports contain the template text, interval, assigned craft and completion status. The CMMS asset hierarchies page covers how the asset tree affects reuse.

Equipment manufacturers. OEM maintenance schedules give recommended intervals and tasks. Facilities adopt them or modify them, so the supplier's local version may differ from the original.

Operator rounds. Walkdown sheets record gauge values, noises, leaks and visual checks along a route. See operator rounds and walkdowns.

Fleet and facility maintenance groups. They hold annual inspection lists, filter and fluid schedules and seasonal procedures, often as spreadsheets or PDFs.

Completed checklists. A signed, dated checklist with as-found and as-left readings is the strongest form, because it records what happened and not only what was planned.

What the data can support, and what it cannot

Checklists give task structure, ordering, constraints and success criteria; they contain no motion, force or perception data. A model that plans from a checklist still requires skills that come from robot data. The page on what SOPs can and cannot teach goes further into that distinction.

Checklists also hide variance. A line reading "inspect belt" may involve two seconds of looking at one site and a tension measurement at another. Without completed records or video, the buyer cannot tell the difference.

Language preparation

Maintenance text is full of shorthand: "chk", "R&R", "brg", "PM-2", site-specific codes. MaintNet is a library of logbook data from aviation, automotive and facilities domains, shipped with abbreviation lists, morphosyntactic lists, termbanks and Python tools for pre-processing and clustering (Akhbardeh et al., 2020). It demonstrates the kind of normalization work required before maintenance text is usable as training or evaluation material. For technician shorthand specifically, see normalizing technician shorthand.

Illustrative record

# Illustrative only - synthetic values
template_id: PM-AHU-QTR-03
asset_class: air_handling_unit
interval: {type: calendar, every: quarter}
craft: hvac_technician
loto_required: true
steps:
  - id: 1, text: "Isolate and lock out unit", links_to: lockout_procedure_L12
  - id: 2, text: "Inspect filter bank; replace if differential pressure above limit"
  - id: 3, text: "Check belt tension and condition"
  - id: 4, text: "Lubricate bearings per schedule"
  - id: 5, text: "Restore, run unit, record supply temperature"
acceptance:
  - {step: 2, check: "dp_reading <= site_limit"}
  - {step: 5, check: "supply_temp within band_B"}
tools: [manometer, tension_gauge, grease_gun]
parts: [filter_type_F, belt_type_V]
completion_record:
  as_found: {dp: 0.9, unit: in_wc}
  as_left: {dp: 0.3, unit: in_wc}
  performed_by: tech_A07 (pseudonymous)
  duration_min: 62

The sample links to a lockout step. Energy-control procedures are a safety issue in every powered task, covered in lockout and tagout procedures.

Ranking scheduled tasks as robot candidates

Checklists let a buyer compare tasks on dimensions that matter for early automation. Useful columns to compute from a template library: how often the task occurs across assets, how many distinct steps it has, how many tools it needs, whether it needs energy isolation, whether it requires measurement and judgment, and how often completed forms record a failed check. Short, frequent, tool-light tasks with a binary acceptance check, such as swapping a standard filter or recording a gauge value, tend to rank high. Tasks needing diagnosis, awkward access or judgment about wear rank lower.

This ranking is a hypothesis about difficulty built from text. Test it against video or a pilot before committing collection budget, and note that the same task title can hide very different physical work across sites.

Quality and acceptance criteria

• Each template tied to an asset class and interval, not floating text.
• Completed records, if included, carry as-found and as-left values and a date.
• Language normalization documented: abbreviation list, mapping rules and error rate on a sampled set.
• Version history of templates, so changes in interval or limit can be dated.
• Split by site or asset family if you will use the data for evaluation.
• Sensitive fields (customer names, building addresses, personnel names) handled according to your agreement.

Limitations and extra preparation

The checklist does not guarantee that work was done. Completed forms are often filled in bulk, signed late or checked off without measurements. Check what fraction of items contain a recorded reading, because a form of unmarked ticks says little. Intervals reflect risk tolerance and warranty terms, not necessarily need.

To use a checklist for physical learning, pair it with demonstrations of the same tasks, ideally with the same step IDs, so a recorded filter replacement can be mapped to line 2. If you plan to commission such recordings, work order histories help identify which tasks actually occur most often.

Permissions and provenance

Checklists can be the supplier's intellectual property, derived from an OEM manual, or authored for a customer under contract. Diligence questions include who wrote the template, whether any text is copied from manufacturer manuals, whether customer sites or equipment are named, and whether the supplier's agreements with clients or unions affect use. Redacting names does not by itself settle who may license the content.

Questions to ask suppliers

1. Are these templates, completed records, or both?
2. How many asset classes and sites do they cover, and over what period?
3. What fraction of completed items has a numeric or free-text reading?
4. Which parts are derived from OEM documents or third-party standards?
5. How are abbreviations and local codes defined, and is there a glossary?
6. Do records link to work orders, parts used and labor time?
7. Were templates revised during the period, and is the history retained?
8. Who can authorize AI-training use, and are there customer or union restrictions?

Discuss your requirements

If you are choosing early robot targets from scheduled maintenance, describe the asset classes, record types and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses. A request does not guarantee a match. The catalog entry for maintenance logs describes the nearest record type, and the entry for field service maintenance records covers service-organization records.

Discuss your humanoid training-data requirements with SourceX

Sources

• MaintNet: A Collaborative Open-Source Library for Predictive Maintenance Language Resources — logbook data, abbreviation lists and pre-processing tools for maintenance text.