Short answer
Scope a manufacturing task for humanoid data by writing down its cycle time, variation, reach envelope, part handling, fixturing and safety constraints before you ask anyone for footage. Those parameters decide which viewpoints, sensors, sites and labels the data needs, and they show when a task is better served by a fixed robot, a different sourcing route or a narrower first pilot.
The buyer's problem
"Assembly" is not a task. A buyer who asks for assembly data gets a mix of seated bench work, line-side kitting, fastening with powered tools and visual checks, each with different physical demands and different evidence in the plant. Suppliers then answer the loosest reading of the request, and the buyer discovers months later that the clips show the wrong reach, the wrong pace or the wrong parts.
Scoping is the work of turning a job title into a bounded task description that a supplier can check against what it holds or could record. It also tells you early whether a task is a sensible target for a humanoid at all, which is a separate question from whether data exists.
Start from a shared task vocabulary
Name the task in terms that both sides can verify. Two public vocabularies help.
Methods-Time Measurement (MTM) is a method standard maintained by the MTM Association for describing, evaluating and designing human work; MTM-style analysis breaks manual operations into basic motions such as reach, grasp, move, position and release (MTM Association). A motion-level breakdown shows where a task is hard: many positions with tight tolerance, or long reaches, or repeated regrasps.
The O*NET database publishes occupational task statements and detailed work activities, plus a file mapping statements to activities (O*NET Resource Center). Its labels are written for occupations rather than robots, so treat them as a naming aid for the request, not as a specification of motion.
Neither vocabulary says anything about what a humanoid can do. They give you words for the task; the rest of this page gives you the parameters.
Parameters that decide the data
| Parameter | What to write down | Why it changes the data |
|---|---|---|
| Cycle time | Target seconds per unit and how much it varies between operators | Short cycles bring motion blur and fast hand motion; long cycles need segmenting and carry more idle time |
| Variation | Part variants, orientations, lot-to-lot differences, changeovers | Variation sets how many distinct situations must be covered, not just how many hours |
| Reach envelope | Heights, distances, two-handed spans, bending and stepping | Compare against the robot's reach before collecting; footage of tasks the body cannot reach teaches little |
| Part handling | Mass, size, fragility, deformability, surface finish | Heavy or compliant parts raise the value of force and contact data |
| Fixturing | Jigs, clamps, trays, part presentation | Controlled presentation simplifies perception; loose presentation is closer to hard cases |
| Tooling | Hand tools, powered tools, gloves | Tool use and gloves affect hand visibility and tracking |
| Quality gate | Pass criteria, inspection step, rework loop | Defines success labels and what failures look like |
| Safety | Guarding, human proximity, lockout, speed limits | Determines whether recording and later deployment are possible |
Fill this table for one task at a time. A task with a different answer in three rows is probably two tasks.
Choosing tasks for a first program
Ranking candidates is a judgment the buyer owns. These checks tend to separate good first targets from poor ones.
- Bounded variation. Can you list the part variants on one page? If not, narrow the task before collecting.
- Visible hands. Is the work done in front of the body, where a head-mounted or fixed camera can see both hands? Work inside enclosures or under assemblies hides what you want to label.
- Observable success. Can an annotator or a downstream station tell from video whether the unit was right?
- Reach fit. Does the task sit inside the body's working volume without climbing or kneeling?
- Pace tolerance. Is the task done at a rate the robot team could plausibly approach, or only at expert line speed?
- Cooperation with machines. Does the operator wait on a machine cycle? Such gaps matter for the plan; see machine tending data.
- Recordability. Can cameras go in without breaking guarding, sight lines or confidentiality rules?
Tasks that fail several checks can still be worth pursuing, but they need a different data mix, such as teleoperated robot demonstrations or instrumented capture, rather than more footage.
What kind of data each parameter points to
Once the task is bounded, map it to the evidence you will need.
Cycle-time and sequence questions are answered by time-and-motion study recordings and standard work and job breakdown sheets, which describe how long steps take and in what order. They do not show how the hand moves.
Motion and hand-object questions are answered by assembly demonstration video, preferably from known viewpoints with tracked hands. Even then, human video contains no robot joint states or commanded actions, so a humanoid still needs robot-specific data for control; see the human-to-humanoid embodiment gap.
Reach and layout questions are answered by workcell layouts and workstation geometry. Safety questions are answered by machine safety constraints for recording.
Illustrative task scope record
# Illustrative only - synthetic values, adapt to your task
task_id: mfg_task_003
name: "load two-part subassembly into test fixture"
cycle_time_target_s: 40
cycle_time_spread: "operators vary; record by operator_A12..operator_A15"
variants: [variant_a, variant_b, variant_c]
reach: "bench height, both hands, no stepping"
part_mass_range: "light to moderate; see part list"
fixturing: "tray with locating pockets"
gloves: "thin nitrile, no heavy gloves"
success_signal: "fixture sensor closes; visual check by next station"
excluded: ["rework loop", "machine fault recovery"]
guarding_notes: "no camera inside guarded zone"
unknowns_to_resolve_in_pilot: ["lighting at second site", "variant_c share"]
The "unknowns" row is deliberate. A scope record should list what the buyer does not know yet; a pilot can answer those cheaply, as described in pilot collection scoping.
What a scope cannot settle
A task scope does not tell you how much data a model needs; that depends on architecture, embodiment, task difficulty and how much robot data you add. It does not prove a humanoid is the right machine for the task, and it does not guarantee that any plant holds suitable recordings. Footage made for other purposes was not designed around your parameters, so expect gaps in viewpoint, labels and variation. Evaluation tasks need separate scoping, held out by site, operator or part as your claim requires.
Questions to ask suppliers
- Which of the task parameters above can you document for the footage or site you offer, and which are unknown?
- How many distinct part variants, operators and shifts appear, and how are they identified?
- What is the recorded cycle time distribution, and was any recording done at demonstration pace rather than production pace?
- Where are the cameras relative to the operator's hands, and can hand visibility be sampled before licensing?
- Are rework loops, faults and interruptions included or edited out?
- Is there a written job breakdown or standard work sheet that matches the recorded steps?
- Which areas could not be recorded because of guarding, confidentiality or customer restrictions?
- Who can grant rights to the footage, including parts or designs belonging to customers? This is not legal advice; ask counsel to review.
Discuss your requirements
Once you have a bounded task record, describe it to SourceX along with modality, volume, geography and permitted uses. SourceX reviews the requirement and follows up about potential sources among US businesses; it checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. For wider context, see the manufacturing data hub and the humanoid training data guide. Related catalog context is on manufacturing quality records.
Discuss your humanoid training-data requirements with SourceX
Sources
- MTM ASSOCIATION e. V. (Methods-Time Measurement) — MTM as a method standard describing human work as basic motions.
- ONET Database (ONET 29.1 per search snapshot): Task Statements and Detailed Work Activities — occupational task statements and detailed work activities as a naming vocabulary.