Short answer
A torque specification dataset links each fastener to a numeric value, unit, tolerance, tightening sequence, lubrication condition and the procedure step that calls for it. It comes from OEM service documentation and service organizations, and needs careful normalization. It can ground task planning and verification, but it contains no motion or measured force, so it cannot teach a robot how to tighten anything.
The buyer's problem
Fastening is a frequent physical operation in repair and assembly, and many fasteners have a specified value attached to them. In a service manual that number sits in a table, a figure callout, a step sentence or a footnote. Extracting it as free text produces errors that matter: a value in foot-pounds read as newton-meters is wrong by a wide margin, and a "torque plus angle" instruction collapsed to a single value loses the second stage.
A robotics team might want this data for three purposes. Task planners need to know which steps involve a controlled tightening. Verification modules need a target against which a measured value is compared. Language models that explain or check procedures need accurate numeric context. All three are text and structure problems, not motion problems. The what SOPs can and cannot teach page explains why the same limit applies to any procedural document.
This page covers the numeric side. For the broader document families, see the maintenance cluster hub and the humanoid data sourcing guide.
What a fastening record should contain
| Field group | Examples | Why it matters |
|---|---|---|
| Identity | asset or model, component, fastener position ID | Joins the value to a physical thing |
| Requirement | value, unit, tolerance, direction (tighten/loosen) | The number itself, in a normalized unit |
| Staging | pass count, angle after snug, final value | Many joints are tightened in stages |
| Sequence | order or pattern across a bolt group | Tightening order affects joint seating |
| Condition | dry or lubricated, thread locker, new or reused fastener | Changes how the same value behaves |
| Tooling | tool type, drive size, adapter notes | Links to tool selection |
| Procedure link | step ID, preceding and following steps | Ties the value to the task graph |
| Provenance | document, revision, effective date | Lets a reviewer check against the source |
Revision and effective date deserve emphasis. A value that was correct for one production range may be superseded, and a dataset without revision metadata cannot be audited.
Where the data lives
OEM technical publications. Civil and military product publications are sometimes authored to the S1000D specification, an XML standard in which procedural and descriptive content is stored as stand-alone data modules, each identified by a unique data module code in a common source database (S1000D). Where a supplier holds source XML, torque values can be extracted from structured elements rather than parsed from layout. See S1000D and ATA documentation for what that means in practice.
PDF and print manuals. Most automotive, industrial and appliance documentation arrives as PDF. Layout analysis datasets that include manuals as a document category, such as DocLayNet with 80,863 manually annotated pages and 11 layout classes (Pfitzmann et al., 2022), show that table and figure detection is a recognized pre-processing task, and torque tables are the sort of content that needs it.
Service organizations. Dealer service departments, fleet shops and maintenance contractors keep their own copies, annotated job cards and local tolerances. These notes show which values technicians use in practice but are not authoritative.
Work orders. Closed repair records sometimes note a torque applied or a torque wrench used. They are sparse, and work order histories are better for frequency and outcomes than for numbers.
What this data can support, and what it cannot
It can support planning and checking: which steps need a controlled tightening, which tool class applies, in what order a bolt group is tightened, and what a measured result should be compared with. It can help build evaluation questions that test whether a model reads a torque table correctly.
It cannot support learning to apply torque. It contains no hand motion, no reaction forces, no tool pose and no information about how a thread feels as it seats. Measured data of that kind comes from instrumented tools or from robot hardware with wrist force sensing. Power-driven fastening has its own considerations in power tool use demonstrations.
It also does not show whether technicians follow the specification. Procedure and practice often differ, and a dataset built from manuals alone cannot say which.
Preparation steps
- Parse documents into tables, steps and figure callouts.
- Extract values with units, tolerances and conditions.
- Normalize units into one system and keep the original string.
- Link each value to a step ID, component and fastener position.
- Validate a sample against the source page by hand.
- Record document revision and effective date per row.
- Remove customer names, serial numbers and internal part numbers if the supplier requires it.
Illustrative record
# Illustrative only - synthetic values
record_id: fx_000412
asset_class: pump_assembly_type_Q
component: housing_cover
fastener_position: bolt_3_of_8
requirement:
value: 40.7
unit: N·m
original_text: "30 lbf·ft"
tolerance: "+/- 5%"
stages: [{pass: 1, value: 20}, {pass: 2, value: 40.7}]
pattern: cross-sequence
condition: {lubricant: light_oil_on_threads, fastener_state: new}
tooling: {tool: click_torque_wrench, drive: 1/2_in}
procedure_link: {step_id: RP-07-14, previous: RP-07-13}
provenance: {document: manual_rev_C, effective: 2022-03-01, page: 118}
review: {checked_by: reviewer_02, status: matches_source}
Reading the numbers: common extraction traps
Several patterns recur in service documents and are worth testing for explicitly. Ranges ("25 to 30") must stay ranges, not collapse to the midpoint. Footnotes that change a value for a specific engine variant or a reused fastener must travel with the row. Torque values given for dry threads differ from those for lubricated threads, so a missing condition field is a data defect, not a minor omission. Units appear as N·m, lbf·ft, lbf·in and kgf·m, sometimes in the same table. Finally, a figure callout such as "item 14" only means something if the dataset keeps the link to the exploded view, covered by exploded views and parts catalogs.
Ask the supplier for an error log from their own validation pass. A supplier that cannot say how often its extraction was wrong has not measured it. For documents built as full procedure sets, service manuals for embodied AI and disassembly and reassembly sequences describe the surrounding structure.
Quality and acceptance criteria
- Sample a set of rows you choose and require every value to match the source page after unit conversion.
- Require the original string to be preserved next to the normalized value.
- Check that staged and angle-based requirements are not collapsed to one number.
- Require revision metadata on every row, with a rule for superseded values.
- Run a held-out test on layouts the extractor has not seen, such as a different manufacturer's table style.
Permissions and provenance
OEM service content is usually authored and copyrighted by the manufacturer, and a service organization holding a copy may only have rights to use it internally. Diligence questions include: who authored the document, under what subscription or license did the supplier obtain it, whether that license allows reuse for AI training, and whether derived structured data counts as a copy. Customer-specific annotations may identify vehicles, equipment owners or sites. Redaction reduces exposure but does not settle who may license the underlying content. This is not legal advice. The OEM service documentation access page covers the access routes.
Questions to ask suppliers
- Is the source structured (XML, database) or scanned and parsed?
- Which manufacturer revisions and model years does each value apply to?
- How were units handled, and is the original text retained?
- Are staged, angle-based and sequence requirements captured?
- What share of rows were checked against the source by a person?
- Who authored the documents, and what license governs their reuse?
- Are site-specific or technician-added values mixed with OEM values, and are they labeled?
- Does the data include fastener position IDs that tie to an exploded view?
Explore sourcing options
If you need fastening requirements tied to specific procedures, describe the equipment families, formats and permitted uses to SourceX. SourceX looks for US businesses that may hold suitable documentation, checks the data and the supplier's licensing permissions, and follows up about potential sources. A request does not guarantee a match. Related catalog entries include internal documentation, and the exploded views and parts catalogs page covers the figures these values refer to.
Explore sourcing and licensing options with SourceX
Sources
- S1000D International specification for technical publications using a common source database — XML data modules and the common source database.
- DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis — layout-analysis data that includes manuals.