Short answer
Assembly demonstration video comes from three places: manufacturers' station cameras and training libraries, time-study recordings, and commissioned head-mounted or multi-view capture. Research corpora show which label layers to request: action segments, hand poses, active objects, step recognition and recorded mistakes. Plant footage adds real pace, gloves and variant mixes but rarely sees hands at the joining moment, cannot show insertion forces, and usually depicts someone's confidential design.
The buyer's problem
Assembly is the most natural video asset for a manufacturing program. Steps are defined, order matters, and plants have many people repeating the same work. The difficulty is that "assembly video" covers products with very different uses. A wide station camera that shows an operator's back and a head-mounted view that shows a clip entering a housing are different purchases, and buyers who do not say which they want tend to receive the first.
This page helps you decide the views and label layers before you ask, and judge what a plant can actually supply. It sits in the manufacturing and assembly hub and the wider humanoid training data guide.
What research corpora suggest asking for
Research teams have recorded assembly of toy vehicles, furniture and construction kits under controlled conditions. They are not substitutes for your line, but their annotation designs are a useful menu.
| Corpus | Capture | Labels | Lesson for your specification |
|---|---|---|---|
| Assembly101 | 8 static and 4 egocentric cameras; 4,321 videos of toy-vehicle assembly and disassembly; unscripted | 100K+ coarse and 1M fine-grained action segments; 18M 3D hand poses; mistakes | Pair fixed and head-mounted views, and label at two granularities (Sener et al., 2022) |
| IKEA ASM | 371 furniture assembly videos; three views; depth | 33 atomic actions; object segmentation; human pose | Fix a closed action vocabulary before annotation starts (Ben-Shabat et al., 2020) |
| HA-ViD | 3,222 multi-view, multi-modality videos in industrial-like scenarios | 96K temporal and 2M spatial labels; human-robot shared annotations | Write labels a robot planner can read as well as a human analyst (Zheng et al., 2023) |
| MECCANO | 20 participants; egocentric; toy motorbike | Temporal action segments; active-object boxes | Ask which object each action touches (Ragusa et al., 2020) |
| IndustReal | Egocentric; a 36-part construction toy car; assembly and maintenance procedures | Procedure step recognition; procedural and execution errors, many only in validation and test splits | Reserve errors for evaluation and track assembly state (Schoonbeek et al., 2023) |
HA-ViD goes further than most: it adds analyses of task collaboration, process efficiency and human intention to the video labels, which is close to what a plant's industrial engineers already produce. See time-and-motion study recordings.
Where assembly video exists in plants
| Source | What it shows | Best for | Watch for |
|---|---|---|---|
| Process-monitoring and quality cameras | Fixed station views | Step timing, station flow, rework context | Hands occluded; short retention; see process-monitoring camera footage |
| Training and onboarding video | Expert demonstrations | Clean reference executions, narration | Staged, edited, sometimes made by outside vendors; see corporate training video archives |
| Time-and-motion study recordings | Work with element timing | Pre-segmented clips | Consultant-owned analyses |
| Quality investigation recordings | Failures and rework | Errors with known causes | Skewed toward problems |
| Visual work instructions | Photos or short clips per step | Step reference imagery | Not continuous video; see visual work instructions |
| Commissioned recording | Head-mounted, wrist or multi-view | Designed views and labels | Site access, safety review, worker consent |
The choice between the first five rows and the last is the usual existing-versus-new decision; existing footage vs commissioned recording gives the procedure. If no existing asset fits, SourceX can scope a commissioned collection of new recordings; the site describes a separate physical-world path for new recordings of hands-on work.
What assembly video can and cannot support
Can support: step and assembly-state recognition; sequence extraction when paired with assembly sequence data; hand-object interaction labels; pace and cycle analysis; mistake detection; language grounding when matched to work instructions; and evaluation scenarios.
Cannot support on its own:
- Insertion and fastening forces. Seating a connector or pressing a clip depends on feel; see snap-fit and press-fit assembly.
- Sub-millimeter alignment. Video at station resolution does not measure it.
- Robot actions. There are no joint states or commanded motions in human video; see the human-to-humanoid embodiment gap.
- Hidden joins. Latches, internal connections and fasteners behind covers are not visible.
What to specify
- Views. Head-mounted or wrist for hands at the joining moment, plus a fixed view of the whole station, synchronized. Fixed cameras alone usually fail the first test.
- Resolution on target. Enough pixels on the part and fastener at working distance, stated as a requirement rather than as a camera model.
- Frame rate and timing. Constant frame rate and real timestamps.
- Unit identity. Serial or lot, product variant and routing step for each clip, so footage links to production records.
- Labels. Step segments at two levels, active objects, hand visibility, assembly state, and errors flagged as natural or staged.
- Conditions. Gloves, PPE, line speed, shift, and operator experience band.
- Splits. Operators, lines and variants reserved for evaluation.
# Illustrative only — synthetic plant, product and values; not an available dataset
package: station_assembly_v0
site: plant_021, line_B, station ST_30
product: housing_assembly_variant_C (customer-owned design: approval attached)
video:
head_mounted: 1080p, constant 30 fps, per-operator device ID
fixed_station: 1080p, constant 30 fps, synchronized by slate at shift start
labels:
step_segments: {coarse: 8 steps per routing, fine: atomic actions, closed vocabulary v2}
active_object: per fine segment
hand_visibility: per second (left, right, both, none)
assembly_state: per step end
errors: {provenance: natural | staged, type, step}
joins: {serial, variant, routing_step, operator_id (pseudonymous), shift}
excluded: [insertion forces, robot actions, customer markings (masked)]
rights: training and internal evaluation; customer approval attached
Quality and acceptance criteria
- Hands at the joining moment. Per step, the share of sampled frames showing hands and the part together, reported per step rather than as one average.
- Step labels match routing. Segments line up with the plant's routing or work instruction on an audited sample.
- Variant coverage. Counts by variant, operator, shift and line, with gaps stated.
- Errors retained. Rework, omissions and corrections are kept and flagged, not trimmed.
- Timing integrity. Constant frame rate, real timestamps and a measured offset to any production event.
- Join rate. The share of clips linked to a unit serial and variant, with unmatched clips delivered separately.
- Permissions. A consent and customer-approval reference for every clip.
Generic video checks are in demonstration video acceptance criteria.
Limitations and extra preparation
Plant footage rarely arrives labeled, so budget for annotation; step segmentation and hand-object interaction annotation describe the work. Hand pose, if you need it, usually requires commissioned capture; see 3D hand pose data. Research corpora report how much they recorded, but no study sets a floor for your task, and volume requirements depend on architecture, task, embodiment and intended use. Assembly of one product at one station says little about another; see high-mix, low-volume assembly.
Questions to ask suppliers
- Which cameras record assembly, from where, and do any see hands at the point where parts join?
- How many products, variants, lines and operators are represented, and over what period?
- Can clips be linked to unit serials, variants and routing steps?
- Are rework, omissions and recoveries kept?
- Is there narration or an associated work instruction for each task?
- Are gloves, PPE or machine guards routinely in the way?
- Whose product designs appear, and has that customer approved use for AI training?
- What notice or consent covered operators, and does it include licensing to an outside company?
- How long is footage retained, and who can export it?
Permissions and provenance
Assembly footage shows product structure, part markings and sometimes unreleased designs. When a plant builds to a customer's design, the customer may need to approve licensing; contract manufacturers as data suppliers and manufacturing data confidentiality set out the questions. Operators appear in every clip; see worker consent and identifiable people. Masking part markings reduces exposure but does not by itself settle permission. This is not legal advice.
Discuss your requirements
If you are specifying assembly video, describe the products, stations, views and labels you need to SourceX. SourceX investigates whether US manufacturers hold suitable footage, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match; SourceX reviews it and follows up about potential sources. The catalog entry for manufacturing quality and inspection records and the manufacturing industry page give wider context.
Discuss your humanoid training-data requirements with SourceX
Sources
- Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — multi-view capture, two label granularities, hand poses and mistakes.
- The IKEA ASM Dataset: Understanding People Assembling Furniture through Actions, Objects and Pose — depth, atomic actions and pose labels.
- HA-ViD: A Human Assembly Video Dataset for Comprehensive Assembly Knowledge Understanding — industrial-like assembly with shared human-robot annotations.
- IndustReal: A Dataset for Procedure Step Recognition Handling Execution Errors in Egocentric Videos in an Industrial-Like Setting — step recognition and held-out errors.
- The MECCANO Dataset: Understanding Human-Object Interactions from Egocentric Videos in an Industrial-like Domain — egocentric assembly with active-object labels.