Short answer
Warehouse onboarding and training materials describe how tasks should be done: step order, safety rules, quality checks, system screens and examples of correct and incorrect work. They can serve as task specifications, language and evaluation items for embodied AI, but they show idealized practice, contain no robot actions or forces, and are often vendor-owned courseware.
Why training content is a task specification
Busy warehouses typically train people for a familiar handful of jobs: receive, put away, pick, pack, load, count, handle returns. Operators often write down how, because new hires turn over and errors cost money. Order picking is described in the literature as the most labour-intensive and costly warehouse activity, with cost estimated at as much as 55% of total warehouse operating expense (de Koster et al., 2007). That pressure may help explain why picking and packing procedures are often among the better documented tasks in a facility.
For a humanoid program, this material is closest to a requirements document. It states what a competent worker does and how supervisors judge it. It is not a demonstration of how a body does it. The limits of written procedures apply in full, and this page focuses on the warehouse-specific forms and how to evaluate them.
Forms of warehouse training material
| Form | Typical content | Most useful for | Common problem |
|---|---|---|---|
| Onboarding slide decks | Role overview, safety rules, quality standards | Task taxonomy, constraints, vocabulary | Generic; little step detail |
| Work instructions / job aids | Numbered steps with photos | Step order, success criteria | Photos are staged and low resolution |
| Training videos | Trainer demonstrations, often scripted | Visual examples of correct handling | Recorded at a slow pace from one angle |
| LMS modules and quizzes | Scenario questions, scored answers | Evaluation items, error categories | Content is vendor-owned |
| Handheld and WMS screen guides | Screen-by-screen workflow | Scan and confirm sequences | Screens expose client or system detail |
| Safety and equipment certification materials | Lifting, forklift and pallet rules | Safety constraints | Covers people, not robots |
For wider coverage of recorded training libraries, see evaluating corporate training video archives as demonstration data. The catalog entry on training materials and LMS content describes the document side.
What the data can support, and what it cannot
Training content can support:
- Task decomposition. A defined list of steps and decision points for picking, packing or receiving, usable as a schema for annotating other data. See the warehouse task taxonomy.
- Language for instruction-following. Phrases, role terms and exception wording that match what workers actually say.
- Evaluation items. Quiz questions and scenarios with scored answers can be turned into procedure-understanding tests for a language or vision-language model.
- Error and exception catalogs. Many modules list typical mistakes such as wrong location, short pick or damaged unit.
- Planner constraints. Safety and quality rules that a task planner must respect.
It cannot support:
- Physical control. Nothing here records joint motion, grasp, force or contact. A humanoid learns manipulation from robot, teleoperated or instrumented data.
- Real pace and variation. Trainers perform slowly and cleanly, so videos underrepresent clutter, occlusion and recovery.
- Site-specific truth. Practice on the floor drifts from the written method, especially under peak load. Compare against engineered labor standards or recorded work where available.
Rights: courseware ownership
Training content often has layers of ownership. The operator may have authored the slides and filmed the videos itself, but modules delivered through an LMS may be licensed from a courseware vendor, and stock footage, music or software screenshots may sit inside them. A 3PL's materials may also show client-specific procedures or branded goods; read client-owned inventory rights in 3PL footage for those questions. Where trainers or actual employees appear on camera, worker consent and identifiable people applies.
Illustrative package: onboarding materials prepared as task specs
# Illustrative only - synthetic structure, not an available dataset
package: wh_onboarding_specs_v0
source: one operator's pick/pack/receive onboarding library (2 sites)
documents:
work_instructions: PDF export + structured steps (JSON)
videos: MP4 with captions and transcript
lms_items: question, options, correct_answer, error_category
structure:
task_id: pick_each | pack_carton | receive_case | putaway
step_id, step_text, tool_or_screen, quality_check, safety_note
links:
video_timecode_to_step_id: yes (manual alignment)
authorship:
operator_authored: true/false per file
third_party_content: flagged per file
excluded: client names, order data, employee names
rights: internal training and evaluation use; no redistribution
The step-to-timecode alignment is the piece that adds value. A video with captions and a matching numbered work instruction is far easier to turn into annotated data than the same two files delivered separately.
Quality and acceptance checks
- Open ten random documents and confirm they are the current version, not archived drafts.
- Check that step numbering is consistent, so a step ID in one file matches the same step in another.
- Confirm that quiz items have the correct answer recorded, along with the reason.
- Verify that videos are not just screen recordings of slides, if you want footage of the physical task.
- Ask what proportion of materials describe tasks your robot would plausibly perform. A large LMS library may contain mostly compliance training.
Limitations and additional preparation
Plan for conversion work. PDF and slide formats need structure extracted; videos need transcripts and step alignment; screens need redaction. Terminology also varies between operators, so mapping step names to a shared taxonomy takes effort. If you plan to evaluate a model on procedure understanding, hold out whole sites or document sets so the test is not drawn from the same material you trained on. Finally, treat training content as the claim of what should happen. To learn what does happen, pair it with recorded work, as in packing demonstrations or time and exception records.
Turning materials into annotation schemas
The most practical use of training content is often indirect: it gives you the label set for other data. A picker onboarding module that lists seven pick exceptions (short stock, damaged item, wrong location, label unreadable, item overweight, multiple items in slot, system mismatch) can become the exception vocabulary for annotating floor video. A packing quality standard becomes the success criteria for packed-carton images.
When evaluating a library, ask whether the same vocabulary is used consistently across modules and across sites. A taxonomy built from inconsistent source terms will carry that inconsistency into labels. Mapping effort is modest for one operator and grows quickly across several, which is a reason to start with a single coherent library before combining sources.
Operators that run training for several clients are covered in 3PL providers as data suppliers.
Questions to ask suppliers
- Who authored each training asset, and which items come from a courseware vendor or stock library?
- Which tasks are covered, and at what level of step detail?
- Are the videos staged by trainers or recorded during live operations?
- Is there a written procedure that matches each video, with step numbers or timecodes?
- Do materials show client names, branded products, system screens or order data?
- How often are materials revised, and can you provide version dates?
- Are LMS quiz items exportable with correct answers and categories?
- Which employees or trainers appear on camera, and what notice or consent applies?
- Do your vendor contracts allow licensing the courseware for AI training?
Discuss your requirements
If you want warehouse task specifications to sit alongside demonstration video, describe the tasks 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; SourceX follows up about potential sources. For the wider map, start at the humanoid training data guide or the warehouse and fulfillment hub. The catalog entry for SOPs and playbooks describes related document types.
Discuss your humanoid training-data requirements with SourceX
Sources
- Design and control of warehouse order picking: A literature review — order picking as the most labour-intensive and costly warehouse activity; cost estimated at up to 55% of operating expense.