Short answer
Small appliance and consumer product assembly mixes screws, clips, snaps, wires and cosmetic parts in short cycles. Multi-view video with step labels, including mistakes, can support step recognition and procedural understanding, as public assembly datasets show. It does not contain robot actions or forces, and factory footage needs permission, labels and often new capture.
Why this segment needs its own page
Consumer product lines differ from heavy industrial assembly in ways that matter for data. Products are small, handled one-handed or two-handed at a bench, and built from many part types: molded housings, springs, gaskets, PCBs, wire leads, labels, and visible cosmetic surfaces that must not be scratched. Models change often as colors, trims and regional variants rotate. Volumes can be high, which means short cycles and fast hands.
This diversity is useful for generalization and awkward for collection. A single line may switch among several variants in a week, each with different step lists. For the general case of assembly video, see assembly demonstration video; for the sequence side, see assembly sequence data. This page covers what is specific to small appliances and consumer goods.
What a public assembly dataset looks like
Assembly101 is a useful reference for what well-designed assembly video includes. It holds 4,321 videos of people assembling and disassembling 101 toy vehicles, recorded with 8 static and 4 egocentric cameras. It provides more than 100,000 coarse and 1 million fine-grained action segments and 18 million 3D hand poses. The recordings are unscripted and include mistakes, and the paper introduces a mistake-detection task (Sener et al., 2022).
Four design choices there are worth copying into a buying specification: multiple synchronized viewpoints, both fixed and first-person; labels at coarse and fine granularity; hand pose; and unscripted behavior with errors left in. These describe one study's data, and your needs depend on your architecture and use. They are also toys, assembled by volunteers, so they do not show factory pace, tooling or cosmetic-quality constraints.
Task anatomy for small product assembly
Typical operations across a small appliance line:
- Part feed and presentation. Housings, subassemblies and loose parts delivered in trays or bins.
- Fastening. Driving screws with a powered driver, placing clips, folding tabs.
- Snap and press fits. Pressing covers, seating buttons, fitting seals. See snap-fit and press-fit assembly.
- Wire and connector routing. Dressing leads, plugging connectors.
- Labeling and marking. Applying rating labels or serial stickers.
- Functional test and visual check. A power-on test, or a cosmetic inspection under light.
- Packout. Moving the finished unit to a carton; see end-of-line packaging data.
Failure modes: cross-threaded or over-tightened screws, pinched wires, clips not fully seated, scratched cosmetic faces, wrong variant parts, missing gaskets. Many failures are invisible in video until the test station, which is why linking test results to the video helps.
What the data can support, and what it cannot
Recorded assembly with step labels can support:
- Step segmentation and recognition, including mistake detection as in the public dataset above.
- Procedural understanding and task ordering for a planner.
- Visual representation learning on parts, tools and hands in context.
- Evaluation protocols for step-level success, with held-out products or stations.
It cannot support:
- Robot control. Hand pose from video is not a robot action. Retargeting to a humanoid hand loses information; see the embodiment gap.
- Force-sensitive steps. Screw torque, snap force and clip seating need force or tool data. For tool-side records, see screw driving and torque controller data.
- Deformable parts and wires. Handling wire leads, gaskets and foam requires contact data that video does not hold.
- Cosmetic judgment. A worker's decision that a surface is acceptable depends on lighting and viewing angle that standard cameras may not reproduce.
Sourcing routes
Existing assets. Consumer goods manufacturers hold work instructions, assembly drawings, training videos, process-monitoring footage and quality records. Training videos are cleanest but often staged. Line footage is realistic but at fixed angles and a fast pace. Check manufacturing data confidentiality because unreleased products are commercially sensitive.
New collection. Because variants and viewpoints drive value, a designed recording is often the stronger option: several fixed cameras, head-mounted views, tracked hands, a step script, and deliberate coverage of variants and common mistakes. This needs careful consent, site access and a decision about how many variants to cover.
Specialist capture. If you need torque, force or robot-frame data, that requires instrumented tools or robot teleoperation.
Illustrative annotation schema
# Illustrative only - synthetic fields, not an available dataset
unit: small_appliance_variant_V03
videos: 3 fixed + 1 head-mounted, synchronized, constant frame rate
segments:
coarse: [prepare, fasten_housing, route_leads, close_cover, label, test]
fine: verb + object + tool, e.g. press(clip_07, housing)
hand_pose: 3D where tracked, confidence per frame
mistakes: type [wrong_part, unseated_clip, scratched_face], timestamp, corrected_flag
context: variant_id, operator_A12, station_id, shift
outcome: test_pass | rework
Quality and acceptance checks
- Review a sample of fine-grained segments by two annotators and measure agreement on boundaries.
- Confirm that mistakes are left in, not edited out.
- Check coverage by variant and operator so one product or person does not dominate.
- Verify that hands are visible in a stated share of frames, and report occlusion by tooling.
- Hold out entire variants or stations for evaluation.
Limitations and additional preparation
Fast cycles produce motion blur and short step durations, so frame rate and shutter settings matter. Camera angles from a line are rarely matched to a robot's head or wrist cameras. Gloves, small parts and tool occlusion reduce hand-tracking quality. Product design confidentiality can restrict what views are permitted, and redaction of logos or screens affects usable detail. Time and money for annotation are substantial for fine-grained labels.
Choosing between existing footage and new capture
Existing training or line footage is cheaper to obtain and reflects real pace. New capture costs more but can match your robot's viewpoint, include head-mounted cameras and hand tracking, and cover the variants you care about. A common compromise is to license existing footage for visual and procedural pretraining, then commission a smaller multi-view recording for the steps you plan to automate. The comparison of existing footage and commissioned recording sets out the trade-offs, and high-mix low-volume assembly discusses variant coverage.
For electronics-heavy products, see also manual electronics assembly.
Questions to ask suppliers
- Which product families and variants are covered, and how many units per variant were recorded?
- Are recordings of normal production or of staged runs?
- How many camera views exist per station, and are they synchronized?
- Are mistakes, rework and recoveries included?
- Do you hold test results that can be linked to units in the video?
- What share of frames shows both hands and the part?
- Are products pre-release or customer-confidential, and what restrictions apply?
- What notice did workers receive, and are faces or tattoos visible?
- Would you support a recording with head-mounted cameras and a step script?
Discuss your requirements
If you need assembly data for small products, describe the product types, cameras 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 broader map, start at the humanoid training data guide or the manufacturing and assembly hub. The site's entry on first-person video of skilled manual work describes the related physical-world path.
Discuss your humanoid training-data requirements with SourceX
Sources
- Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — multi-view assembly video with action segments, hand poses and mistakes.