Short answer
An electronics assembly dataset for humanoid work is usually video of people placing, inserting and fastening small components, labeled by step and by acceptance outcome. It can support step understanding and outcome labeling. It cannot train hand control by itself, because footage carries no robot actions or contact forces. Small parts and static-control rules also make capture design critical.
The buyer's problem
Electronics assembly is a poor fit for the camera angles that much workplace footage uses. The parts are small, the hands cover them, and the decisive moment, a lead entering a hole or a connector seating, happens in a few square centimeters. A ceiling camera that shows a worker "doing assembly" tells you the task was performed. It rarely shows whether the part went in correctly.
A buyer therefore has to decide what the dataset is for before looking at suppliers. Step recognition, outcome classification and scene understanding tolerate ordinary viewpoints. Anything closer to manipulation needs close views, known calibration and, for control, robot-side data. The wider logic of what human video can and cannot do is set out in what human demonstration video can and cannot do for robot learning.
Task anatomy
Manual electronics assembly covers several different hand skills that are often lumped together:
- Through-hole insertion: picking a component, forming or checking its leads, aligning the leads to holes on a board and pressing it down, sometimes with a board fixture or a magnifier.
- Mechanical fastening: starting and driving small screws, placing standoffs, clipping a board into a housing.
- Connector and cable work: seating a board-to-board or cable connector, routing a flex cable, closing a latch. See connector mating and precision insertion and wire harness assembly.
- Box build: assembling boards, harnesses, panels and enclosures into a finished unit, applying labels, and handing it to test. The test step is covered in end-of-line test station loading.
Typical failure modes include a reversed polarized part, a bent lead, a component seated crooked, a dropped screw, a cable pinched at closure and a missed step in the sequence. Those errors are the useful negatives in a dataset, and a supplier that edits them out has removed the most informative material.
Why capture is harder here
Electronics floors often have their own static-control and cleanliness rules, and this page does not state what they are. The consequence for a recording project is practical: equipment brought onto the floor, tripods, cables and head-mounted cameras, may need approval from the site, and what is permitted differs by facility. Treat this as a question for each supplier rather than an assumption. The special case of controlled rooms is treated in recording assembly in cleanrooms.
Small parts also create a labeling problem. Annotators need to see the part to mark insertion events, so footage shot at wide angles yields coarse labels. Some stations use magnifiers or microscopes, which means the operator's view differs from any camera on the bench. If you want the operator's view, ask whether the magnifier output can be recorded.
What data helps each part
| Need | Data that helps | Typical source |
|---|---|---|
| Step order and completeness | Overhead and oblique video with step segments | Existing training footage or commissioned recording |
| Small-part events | Close-up or macro view of the work area | Commissioned recording |
| Part identity | Reference images and part numbers per station | Supplier part records |
| Acceptable or not | Inspector verdicts tied to each unit | Quality records |
| Hand motion | 3D hand tracking or handheld-device capture | Commissioned recording |
How parts arrive at the station, in trays, reels or bags, affects both the footage and the labels; see parts presentation and identification.
Acceptance labels deserve care. IPC-A-610 is described in secondary sources as setting visual acceptance criteria for electronic assemblies, covering solder joints, placement, cleanliness, coating and marking, across three product classes, with the 2024 revision designated J (IPC). A supplier may label outcomes against that standard or against internal criteria. Ask which, and which class applies to the product; the label means different things under different classes. Do not infer from a label alone that the unit met any particular requirement, and treat the standard's details as something to obtain from IPC, not from a data page.
Large assembly video datasets show how this kind of material is organized. HA-ViD, for instance, reports 3,222 multi-view, multi-modality videos with 96K temporal labels and 2M spatial labels, with annotations shared between human and robot views (Zheng et al., 2023). It covers industrial-like scenarios, not electronics, so use it as a reference for label structure only.
Illustrative package
# Illustrative only - synthetic values, not an available dataset
package: electronics_hand_assembly_v0
site: site_044 (synthetic)
station_types: [through_hole_insertion, screw_fastening, cable_seating]
streams:
overhead_video: 1080p, 30 fps
close_view_video: bench-mounted, fixed focus on work area, 60 fps
operator_view: optional, only if site rules allow head-mounted capture
labels:
step_segments: start/end per step, step_id from the station instruction
part_events: insertion_started, seated, rejected_by_operator
unit_outcome: accept / rework, plus acceptance criteria named by the supplier
errors: reversed_part, bent_lead, crooked_seat, dropped_screw
metadata: board_variant board_B7, operator_id operator_A12 (pseudonymous), shift_label
sync: shared clock or visible sync event per recording
excluded: robot actions, joint states, contact forces
rights: training and internal evaluation; customer designs excluded or licensed separately
For planning only, suppose a pilot covers 3 board variants, 4 operators and 6 repetitions each: 72 recordings. These inputs are arbitrary, not minimums; real requirements depend on architecture, task and embodiment.
Limits: what this data cannot support
Video of people assembling boards does not contain robot joint states, end-effector commands or insertion forces. A humanoid manipulation policy needs robot or teleoperation data, or handheld-device capture in the robot's action space; footage, written procedures, quality records and recognition or pose labels alone do not train control. See the human-to-humanoid embodiment gap for why a human hand's technique does not transfer directly.
Other limits follow from the task:
- Occlusion. Fingers and tweezers hide the contact point in many views.
- Scale. Components too small for the camera's resolution cannot be labeled reliably.
- Single-site bias. One product family and one fixture design say little about another.
- Happy-path bias. Footage of finished, accepted units under-represents errors.
Permissions and provenance
Board layouts, bills of materials and part numbers visible in a frame may belong to the manufacturer's customers. Operators are identifiable by face, hands, badges and voice. Redaction or removing identifiers does not settle permissions, and blurring can also remove the detail a model needs. Ask who can grant rights, what notice and consent operators received, and whether customer contracts restrict recording; see worker consent and identifiable people and manufacturing data confidentiality. This is not legal advice.
Keep the purpose clear too. A dataset used to evaluate how a robot perceives assembly states is a different use from watching how fast an individual operator works, and consent and licensing should say which one is intended.
Questions to ask suppliers
- Which operations are covered, and are they recorded at a distance that shows the part-level event?
- Do rules on static control or cleanliness limit where cameras can sit, and how did that shape the footage?
- Which acceptance criteria and product class label the outcomes, and who judged each unit?
- Are errors, rework and corrections kept in, or edited out?
- How many board variants, fixtures and operators appear?
- Can recordings be linked to unit serials and to the written station instruction?
- What do the images show of customer designs, and what has been excluded?
- Can the supplier commission new recordings with chosen viewpoints and calibration?
Discuss your requirements
If you are weighing existing footage against a purpose-built recording for small-part assembly, describe the operations, viewpoints and permitted uses to SourceX. 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 license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. For wider context, see the humanoid training data guide and the manufacturing hub; the catalog page on CAD and PCB engineering files covers the design side.
Describe the workplace demonstration data your robotics project needs
Sources
- IPC-A-610 Acceptability of Electronic Assemblies — visual acceptance criteria and product classes.
- HA-ViD: A Human Assembly Video Dataset for Comprehensive Assembly Knowledge Understanding — multi-view assembly video and label structure, as a reference only.