Short answer
Operator rounds and walkdown records log a fixed route through a plant or building: stops, readings, observations and exceptions. They suit mobile inspection robots because the route, the checkpoints and the pass or fail rules are explicit. Plants, utilities and facilities hold them as forms, logs and photos. They show where to look and what to record, not how to manipulate equipment.
Why rounds records suit mobile robots
An operator round is a structured walk: start at a station, visit a list of checkpoints in an agreed order, read gauges, listen for noise, look for leaks, check lamp states, and write down anything abnormal. The route, the stops and the expected condition at each stop are already defined. Those are the pieces a mobile robot, humanoid or otherwise, needs to plan a patrol and report a finding.
Walkdowns are a related practice: a deliberate inspection of an area or system, often before work, startup or an audit, with a checklist rather than a daily route. Both produce records that tie observations to places and equipment. This guide covers where those records live, what they can train and what they cannot. For tasks beyond observation, see facility maintenance tasks. It belongs to the maintenance cluster of the humanoid data guide.
What a round record contains
| Element | Examples | Value for a robot |
|---|---|---|
| Route and stop list | Ordered checkpoints, area codes, floor or yard locations | Patrol planning; stop ordering |
| Checkpoint definition | Asset, what to check, acceptable range or state | Task specification at each stop |
| Readings | Pressure, temperature, level, counter values, lamp states | Reading targets; value ranges |
| Observation text | "slight seep at flange", "unusual vibration" | Language of abnormality |
| Photos | Gauges, leaks, labels, housekeeping issues | Perception examples with context |
| Exceptions and handoffs | Out-of-range values, work request raised | Escalation rules |
| Timing | Start, end, time at each stop | Realistic route durations |
Plants, utilities and facilities teams hold these as paper sheets, handheld-device exports and entries in maintenance systems. The short, abbreviated shape of such text resembles the logbook data in MaintNet, which collects logbook data from aviation, automotive and facilities domains and provides abbreviation lists and pre-processing tools (Akhbardeh et al., 2020). Observation notes from rounds usually need the same normalization; see normalizing technician shorthand.
What the data can support
- Route and checkpoint modeling. Ordered stops with asset links support patrol planning and checkpoint definitions.
- Reading and observation language. Text can support models that describe abnormal states and decide when a reading needs escalation.
- Anomaly examples. Photos paired with exceptions give labeled examples of conditions an inspection robot should flag. Coverage is skewed toward normal conditions, so rare faults are scarce.
- Evaluation. Records with later outcomes, such as a work request created after a flagged observation, allow evaluation of whether a system flags what a human flagged.
For gauge reading from images, see analog gauge and meter reading. For checklist-style tasks, see preventive maintenance checklists.
What this data cannot support
- Manipulation and control. A round record contains no hand motion, valve turning, door opening or tool use. A humanoid acting on a finding needs robot-specific observations and actions.
- Robot navigation data. Rounds are written by people. They say little about door widths, stairs, floor surfaces, lighting, radio coverage or elevator use, which determine whether a robot can follow the route.
- Sensory judgment. Operators note smell, sound and feel. Records may capture a word for it, not a measurable signal.
- Hidden decision rules. Tolerances and thresholds are often tacit or written elsewhere. Check whether acceptable ranges are in the data.
- Safety-critical areas. Hazardous areas have access rules; see recording in hazardous spaces.
Existing records document what was written, not what the operator saw. A missing entry can mean a normal state or a skipped stop.
Illustrative round record
# Illustrative only - synthetic values, not an available dataset
round_id: rnd_2026_0412_A
site: site_031
route: {name: pump_house_morning, version: 7, stops: 14}
operator: operator_A12
stops:
- seq: 3
asset_id: asset_0207
check: "discharge pressure"
expected: {min: 4.0, max: 5.5, unit: bar}
reading: 3.6
status: out_of_range
note: "dropping since last round"
photo: img_44021
action: {work_request: wr_7712, priority: medium}
- seq: 4
asset_id: asset_0208
check: "seal condition"
reading: null
status: normal
timing: {start: "06:02", end: "06:49"}
Convert units, normalize asset IDs and decide how to treat null readings before use. A perception team will care mainly about photos with context; a planning team about the route and expected ranges.
Acquiring the records
Existing records are the usual route: export the round history for a site, with route definitions and photo links, then license it. Records alone seldom show geometry. If you need the walk itself, such as video from a head or chest camera as an operator completes a route, that is a new collection, designed with the site's consent and safety rules. Existing footage versus commissioned recording sets out the tradeoffs.
Preparation a buyer should budget for
- Link to assets. Join stops to a facility's asset hierarchy so a reading can be tied to a type of equipment.
- Redact. Photos can show names, badges, screens, security features and customer information. Redaction is a planning step; whether the site may license the images is a separate question.
- Version routes. Routes change. Keep route version with each round or the same stop will look inconsistent.
- Check handwriting and scanning quality for paper sheets before agreeing on volumes.
Turning rounds into an evaluation
Rounds records are well suited to scoring a patrol system because they come with a human answer for each stop. A reasonable evaluation design uses held-out sites, or held-out time periods, and compares three things: whether the system visited every stop, whether its reading matches the recorded value within a tolerance you set, and whether it flagged the stops the operator flagged.
Two cautions apply. Operators sometimes record a reading from memory or copy the last value, so a recorded value is not always independent ground truth. And rounds with few abnormal stops give an evaluation that rewards a system for saying everything is normal. Ask for the abnormal share, and sample enough exceptions to test them. Text-heavy notes can be cleaned with the methods described for maintenance logbook text, but keep the original wording for audit.
Questions to ask suppliers
- How many sites, routes and distinct stop types do the records cover?
- Are routes versioned, and are checkpoint ranges stored with each stop?
- Do records include photos, and how were they taken (handheld device, fixed camera)?
- Are skipped stops, interrupted rounds and blank entries distinguishable from normal ones?
- Can out-of-range readings be linked to the work requests or repairs that followed?
- Which records are handwritten, and what scan or transcription quality should be expected?
- Do any areas covered have security, safety or regulatory restrictions on images or data use?
- Who owns the data: the site operator, a contractor or a software vendor that hosts the records?
- Are operators identifiable in notes or images, and what notice were they given?
Discuss your requirements
If your project needs round or walkdown data for a defined type of site, describe the facility type, record format 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, and nothing is contracted until a supplier agrees. The catalog entry for inspection reports describes a related record type.
Explore sourcing and licensing options with SourceX
Sources
- MaintNet: A Collaborative Open-Source Library for Predictive Maintenance Language Resources — Published maintenance logbook language resource showing how short observation text appears in facilities and other domains.