Short answer
Controlled-language procedures, such as those written to ASD-STE100 or packaged as S1000D data modules, restrict vocabulary and structure, which tends to reduce ambiguity and make steps easier to extract. That is a hypothesis to test on your own pipeline, not a guarantee. The text still holds no motion or force information, and its regularity may not match how real workers talk.
What "controlled" means in practice
Most procedures are written by subject experts in their own voice: long sentences, synonyms, implied subjects and local shorthand. Controlled languages replace that with rules. ASD-STE100 describes itself as a controlled natural language and an international standard for writing technical documentation, owned by ASD in Brussels. Its Issue 9, dated 15 January 2025, revises wording in 31 of its 53 writing rules and updates 555 dictionary entries without adding rules (ASD-STE100). In practice, controlled languages of this kind typically combine a restricted vocabulary, in which approved words carry defined meanings, with writing rules that shape sentences and steps; check the details against the issue a supplier follows. The official site notes that an official copy is available on request, so a supplier should be able to say which issue its text follows.
S1000D is different in kind. It is an XML-based specification for producing and managing technical publications, in which data modules are stand-alone units of descriptive, procedural or operational data held in a common source database and identified by a unique data module code (S1000D). Where STE governs wording, S1000D governs structure and reuse. An aerospace maintenance publication may use both.
Why they might parse better
The argument is mechanical, and each part can be tested:
- Lower lexical ambiguity. If a word has one approved meaning, a model or parser has fewer readings to resolve.
- Predictable step form. Consistent sentence patterns can make steps, conditions and warnings easier to segment.
- Explicit structure. Data modules hold procedural content separately from descriptive content, and a supplier's schema may type elements such as warnings and cautions, which would map well to a step schema.
- Stable identifiers. Unique codes let you track reuse and revision across publications.
None of this is a demonstrated improvement for a humanoid pipeline. It is a reasonable expectation that you can measure, and the measurement is likely to cost less than any premium for seeking out this material.
A test to run before you pay for it
Build a small paired sample: a handful of tasks documented in controlled language and the same or comparable tasks in free-form worker text. Run your extraction pipeline on both and compare. Validating LLM-extracted steps describes how to score the output.
The table is Illustrative: metrics are suggestions and the thresholds are yours to set.
| Measure | How to compute | What a gap would mean |
|---|---|---|
| Step segmentation agreement | Share of steps matching a human-segmented reference | Controlled text segments more cleanly, or not |
| Condition extraction | Share of "if/when" clauses captured with their action | Warnings and branches parse reliably, or not |
| Entity resolution | Share of tool and part mentions resolved to one canonical item | Vocabulary control helps naming, or not |
| Warning typing | Share of warnings correctly typed | Structure carries over to safety content |
| Downstream plan validity | Share of plans accepted by a checker | Whether cleaner parsing changes what you need |
If the gap is small, the premium is not justified. If it is large, ask whether the benefit is from the language rules or merely from the structure; S1000D structure may do most of the work without STE wording, and the reverse can also be true.
Illustrative before-and-after
The example is Illustrative and invented; it does not quote any standard or publication.
Free-form: "Once the panel's loosened off, get it out and put it somewhere safe, then check that the seal looks OK before you go further."
Controlled style: "1. Remove the panel. 2. Put the panel in the tray. 3. Examine the seal. 4. If the seal is damaged, replace the seal. Continue only when the seal is correct."
The second version is easier to segment and to convert to a task graph, and it is also less informative: "loosened off", "somewhere safe" and "looks OK" have gone, and with them some of the tacit cues a worker would use. That tradeoff is the reason to read natural worker instructions alongside this page. A system that grounds only on controlled text may stumble when a human speaks to it in the first form.
Where controlled-language sources sit
STE is closely associated with aerospace documentation, and some industrial documentation teams outside aerospace also adopt it. S1000D is used for civil and military products. Such sources tend to live in publishing and technical-documentation departments, often as structured exports; see structured authoring exports for what an export looks like. Illustrated parts and procedures are often tied together, covered in illustrated assembly manuals.
Likely suppliers include aerospace and defense contractors, industrial equipment makers and large maintenance organizations. Their publications are often customer deliverables, which affects rights.
What controlled-language text cannot support
Cleaner text is still text. It contains no joint states, trajectories, forces or perception data, and cannot train a humanoid's physical control; that requires robot or teleoperation data, possibly with tracked human demonstration. Controlled language also cannot show whether a procedure is correct or followed, and it can strip the context a human would infer.
Further caveats:
- Compliance is variable. A document labeled as STE may not follow every rule; check by sampling.
- Domain narrowness. Maintenance and assembly are likely to dominate; do not expect coverage of retail, food or care tasks.
- Mismatch with spoken instruction. Robots that take spoken commands face free-form language, which this data does not represent. See rewriting human procedures for robots for the reverse approach.
- Overfitting to style. A model grounded only on one house style may be brittle outside it.
Permissions and provenance
Technical publications are commonly customer deliverables, and the maker, the customer and a documentation contractor can each claim rights. The STE standard is itself owned by ASD; using its rules to write text is a different matter from copying the standard's own content, and your supplier should be clear about that. Data modules may embed manufacturer part numbers, export-controlled technical data or customer markings. Ask who owns the content and whether any of it is subject to distribution restrictions. This is not legal advice.
Questions to ask suppliers
- Which controlled language and which issue was used, and how is compliance checked?
- Is the text authored in STE, or translated or edited into it afterward?
- Is the source a publishing system export, such as S1000D data modules, or rendered PDFs?
- Are warnings, cautions and conditions typed in the source, and does the export retain the types?
- What share of the content is procedural versus descriptive?
- Who owns the publications: the supplier, a customer or the equipment maker?
- Is any content subject to export-control or customer-marking restrictions?
- Can a sample of five modules be provided in both source and rendered form?
If you want this class of source investigated, describe the sectors, document types and structure you need in a buyer request.
Discuss your requirements
Tell SourceX what you are trying to extract and from which kinds of publications. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match, and conversion work is scoped case by case. For wider context, see the procedures hub, the humanoid training-data section and the procedure step schema.
Explore sourcing and licensing options with SourceX.
Sources
- ASD-STE100 Simplified Technical English — Issue 9 (15 January 2025), controlled natural language standard.
- S1000D International specification for technical publications using a common source database — XML data modules held in a common source database.