Short answer
Manufacturing robot training data records how parts are presented, assembled, fastened, tended and checked on real lines, through assembly video, standard work, routings, time studies, process records and part data. It supports step and assembly-state recognition, sequence planning, part identification and cycle-time analysis. It does not capture insertion forces or fine alignment, and designs and customer information in it are usually confidential.
Scope of this cluster
Humanoid deployments in manufacturing tend to target stations that were built for people: line-side kitting, harness routing, light assembly, machine tending, packaging and rework. Those stations already generate documentation and records, because plants plan work in detail. A station has a routing step, a standard work sheet, a cycle time, a bill of materials and often a camera. This cluster covers how to turn that plant-floor record into data a humanoid program can use, and where it falls short.
It is written for applied deployment teams scoping a line, and it sits mainly in the demonstrations, procedures and evaluation layers of the humanoid training data guide. Generic manipulation skills live in workplace task demonstrations; tool use lives in maintenance and tool use; this cluster owns process-level assembly and station work.
What plant data looks like
| Asset | What it contains | Typical holder |
|---|---|---|
| Assembly and process-monitoring video | Station views of operators at work | Manufacturing engineering, quality |
| Time-and-motion study recordings | Video with element-level timing | Industrial engineering, consultancies |
| Standard work and job breakdown sheets | Ordered elements, key points, reasons | Lean and operations teams |
| BOMs, routings and work instructions | Parts, operations, sequence, parameters | Engineering and planning |
| MES and torque-controller records | Per-unit events, results and fastening curves | Production systems |
| PLM, ERP and CAD part data | Part identity, geometry, revisions | Engineering, often customer-controlled |
| Existing robot programs and cell logs | Paths and I/O for industrial robots | Automation teams and integrators |
What it can support, and what it cannot
Assembly is well covered in video research, and those datasets show what good assembly data contains. Assembly101 recorded 4,321 videos of people assembling and disassembling toy vehicles from 8 static and 4 egocentric cameras, with 18 million 3D hand poses and unscripted mistakes (Sener et al., 2022). IKEA ASM recorded 371 furniture assemblies from three views with depth, 33 atomic actions and pose annotations (Ben-Shabat et al., 2020).
HA-ViD targets industrial-like assembly with 3,222 multi-view videos and human-robot shared annotations (Zheng et al., 2023). MECCANO adds egocentric video of industrial-like assembly with active-object labels, and IndustReal adds procedural and execution errors (Ragusa et al., 2020; Schoonbeek et al., 2023).
Note what those corpora assemble: toy vehicles, furniture, construction toys and lab assembly setups. Real lines add gloves and PPE, production pace, variant mixes, cramped fixtures and confidential parts. Plant data brings that realism, which research sets lack; research sets bring calibration, multi-view coverage and labels, which plant archives lack.
| Plant data supports | Plant data cannot support alone |
|---|---|
| Step recognition and assembly-state detection | Insertion, snap-fit and press-fit forces |
| Sequence planning from routings and standard work | Sub-millimeter alignment and compliance |
| Part identification from part masters and station images | Robot actions for a humanoid's arms and hands |
| Cycle-time and pace analysis from time studies | Behavior of deformable parts such as hoses and harnesses under handling |
| Error and rework examples for evaluation | Safety validation of a humanoid near machines |
Contact-rich steps need robot or instrumented data; snap-fit and press-fit assembly explains when visual data falls short. Industrial robot programs are a different embodiment: existing industrial robot programs and cell logs covers what they can contribute.
Safety shapes both deployment and collection. The industrial robot safety standard ISO 10218-1 was revised in February 2025 (ISO, 2025), and ISO/TS 15066 adds requirements for collaborative operation alongside people (ISO, 2016). Neither is a data standard or written specifically for humanoids, but plant safety rules built on them can limit which stations are candidates and where a recording team may work near running machines.
Sourcing routes
Manufacturers. Plants hold process-monitoring footage, time studies, standard work, routings and MES records. The richest join is video aligned to MES events and torque-controller results, so each clip carries a unit, a step and an outcome. See assembly demonstration video and MES records aligned to video.
Contract manufacturers. They run many products for many customers, which gives variety, but the designs, work instructions and often the footage belong to those customers. Read contract manufacturers as data suppliers before scoping.
Commissioned recording. Station cameras rarely see hands clearly. Head or wrist capture at a cooperating plant gives hand detail and step structure, but needs safety review, worker consent and production sign-off. If no existing asset fits, SourceX can scope a commissioned collection of new recordings.
Specialist capture. Insertion forces and torques come from instrumented rigs or robot data. SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation and engage specialist capture partners, scoped case by case. See force/torque data for contact-rich tasks.
Confidentiality runs through all four routes: designs, process parameters, supplier names and customer identities. See manufacturing data confidentiality.
How to evaluate manufacturing supply
- Ownership. Whose product is being built, and does that customer need to approve use of the footage or documents?
- Visibility. Whether station cameras show hands and parts at the moment of joining, or only the operator's back.
- Joinability. Whether video can be linked to routing step, unit serial and MES result.
- Variant coverage. How many product variants and changeovers appear; see high-mix, low-volume assembly.
- Errors kept. Whether rework, scrap and stoppages are retained in the record.
- Conditions. Gloves, lighting, line speed and shift differences.
- Export and IP flags. Whether any parts or drawings carry export-control or customer restrictions.
Illustrative station-level data inventory
| Station (synthetic) | Existing asset | Main gap | Likely route |
|---|---|---|---|
ST_10 line-side kitting | Overhead camera, pick lists | Hand detail | Commissioned head-camera capture |
ST_20 harness routing | Visual work instruction, time-study video | Clip insertion force | Specialist or robot data |
ST_30 screw-driving | Torque-controller result per screw | No video joined | Align controller logs to station video |
ST_40 end-of-line packing | Fixed camera, packaging specification | Labels | Annotation on existing footage |
ST_50 test-station loading | None | Everything | Commissioned recording |
An inventory like this, built with the plant's engineers, is a practical way to see which stations are worth licensing and which need new capture. Scoping manufacturing tasks for humanoid data covers the cycle-time, variation and reach questions to ask per station.
Key pages in this cluster
| Group | Page | What it answers |
|---|---|---|
| Scoping | Scoping manufacturing tasks | Cycle time, variation and reach per station |
| Scoping | High-mix, low-volume assembly | Data for many variants in small runs |
| Scoping | Snap-fit and press-fit assembly | When visual data falls short |
| Sources | Assembly demonstration video | Sourcing assembly footage |
| Sources | Time-and-motion study recordings | Video with element timing |
| Sources | Contract manufacturers as suppliers | Variety with customer-owned IP |
| Sources | Assembly sequence data | BOMs, routings and work instructions |
| Sources | Standard work and job breakdown sheets | Element-level work documentation |
| Sources | Process-monitoring camera footage | Evaluating existing plant cameras |
| Sources | MES records aligned to video | Per-unit events as labels |
| Sources | PLM and ERP part master data | Part identity for recognition |
| Sources | Screw-driving and torque controller data | Fastening results per screw |
| Sources | Existing industrial robot programs | What cell logs can contribute |
| Tasks | Wire harness assembly | Routing and connecting deformable harnesses |
| Tasks | Machine tending | Loading and unloading CNC and presses |
| Tasks | Connector mating and precision insertion | Tight-tolerance joining |
| Tasks | Line-side kitting and replenishment | Material presentation to stations |
| Tasks | Consumer product assembly | Small appliance and product assembly |
| Tasks | End-of-line packaging | Packing finished goods |
| Tasks | Rework and repair operations | Correcting defective units |
| Tasks | Fixturing and part orientation | Placing parts into fixtures |
| Rights | Manufacturing data confidentiality | Designs, processes and customers |
Neighboring clusters
Assembly verification and defect images are covered in inspection and failure data, including assembly completeness verification. Written work instructions in general are in SOPs and procedural data; fasteners and small parts in product and object data; machinery upkeep in maintenance and tool use. For SourceX catalog context, see the manufacturing industry page, manufacturing quality and inspection records and CAD and PCB engineering files.
Discuss your requirements
List the stations or tasks, product types, the assets you want (video, standard work, MES or part data) and the uses you need licensed. Describe the plant data to SourceX: it investigates whether US manufacturers hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. SourceX reviews each requirement and follows up about potential sources; a request does not guarantee a match.
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 hand poses and mistakes.
- The IKEA ASM Dataset: Understanding People Assembling Furniture through Actions, Objects and Pose — furniture assembly with depth, actions and pose.
- HA-ViD: A Human Assembly Video Dataset for Comprehensive Assembly Knowledge Understanding — industrial-like assembly with human-robot shared annotations.
- The MECCANO Dataset: Understanding Human-Object Interactions from Egocentric Videos in an Industrial-like Domain — egocentric industrial-like assembly.
- IndustReal: A Dataset for Procedure Step Recognition Handling Execution Errors in Egocentric Videos — procedure step recognition with errors.
- ISO 10218-1:2025 Robotics — Safety requirements — Part 1: Industrial robots — industrial robot safety standard.
- ISO/TS 15066:2016 Robots and robotic devices — Collaborative robots — collaborative operation requirements.