Furniture Assembly Data

The buyer's problem

Furniture is the task people reach for when they want a controlled, widely understood assembly benchmark. The parts are large enough to see, the steps are discrete, and a printed manual defines the intended order. That makes it a convenient testbed, and it also creates a trap: a model that performs well on public flat-pack video has learned something about a narrow, well-lit, manual-driven task, not about assembly in general.

This page helps a buyer decide whether public data is enough, where commercial supply adds something, and what the pairing of video with manuals does and does not buy. For the generic mechanics of ordering, see assembly sequence data; for related household-scale products, see consumer product assembly.

What public research already covers

The IKEA ASM dataset contains 371 furniture assembly videos recorded from three camera views, roughly 3 million frames in total, with depth, 33 atomic action classes, object segmentation and human pose annotations (Ben-Shabat et al., 2020). It is a human-activity dataset: its labels describe what people do and where their bodies are.

IKEA-Manual works from the other direction. It pairs 102 IKEA objects with their assembly manuals and annotates decomposed assembly parts, assembly plans, manual segmentation and 2D-3D correspondence (Wang et al., 2022). Its subject is the link between the drawn instruction and the 3D shape, not footage of hands.

Together they show the two halves a furniture buyer usually wants: how people carry out the steps, and how a written instruction maps to parts. Neither provides robot joint states or commanded actions, so both inform perception, planning and instruction following, with control handled elsewhere. Check each dataset's license yourself before any commercial use, since research datasets often carry their own terms.

Task anatomy

Flat-pack assembly follows a recognizable arc:

• Inventory and identify: laying out panels, dowels, cam locks, screws and a hex key, matching them to the manual's part drawings.
• Orient and align: deciding which face and which edge go together, often the step where errors occur; see fixturing and part orientation.
• Join: inserting dowels, tightening a cam lock, driving screws, tapping with a mallet.
• Reposition: flipping or standing a partial assembly, sometimes needing two people or a stable support.
• Check: confirming the frame is square and the doors or drawers move.

Common failures are a panel mounted backwards, hardware used in the wrong hole, a skipped step discovered several steps later, and a part dropped or damaged while holding the assembly. Large panels bring a physical dimension that small-part tasks lack: bracing, two-handed carrying and stability while torquing.

Where commercial supply adds value

Furniture makers and assembly services may hold training videos, installation recordings or customer-service footage for products that no public dataset includes. Assembly services see variation that a lab session omits: uneven floors, mixed hardware kits, damaged parts and customers' unusual constraints. A manufacturer's own technicians may also assemble prototype or returned units in ways a consumer never would.

Whether existing recordings or a new session fit better is the question taken up in existing workplace footage vs commissioned demonstration recording. For furniture, existing footage is often instructional and scripted, which helps with step labels but under-represents mistakes.

Pairing video with manuals

A manual provides ordered steps, part names and a drawn target state. Pairing it with video lets you align each drawn step to a time span, check whether the person followed the intended order, and derive language-grounded supervision. The pairing is only as useful as its alignment: a manual step may cover several physical actions, and people frequently deviate. Ask whether alignments were annotated, by whom, and how disagreements were recorded.

Manuals are also documents with authors. They are protected content in the hands of the manufacturer and are not the same as a procedure you can freely copy. For the broader limits of written procedures, see what SOPs can and cannot teach a humanoid and instruction-paired demonstration datasets.

Illustrative paired-recording package

# Illustrative only - synthetic values, not an available dataset
package: furniture_assembly_paired_v0
products: [bookcase_type_A, wardrobe_type_B]   # synthetic names
recordings:
  views: [front_fixed, side_fixed, head_mounted]
  depth: optional, one view
manual:
  file: product manual as supplied by the rights holder
  alignment: manual_step_id -> start/end time per recording
labels:
  atomic_actions: from a stated action vocabulary
  parts: panel/hardware ids tied to the manual's part list
  errors: wrong_orientation, wrong_hardware, skipped_step, dropped_part
metadata: assembler_id assembler_A07 (pseudonymous), experience_level, floor_setting
excluded: robot actions, joint states, contact forces
rights: training and evaluation use; manual and product design rights stated separately

Limits: what this data cannot support

Furniture video and manuals do not teach a humanoid to hold a panel, align a dowel or tighten a cam lock. Control needs robot or teleoperation data, or handheld-device capture in the robot's action space, not footage, manuals or pose labels alone. Human pose annotations describe a person's body, which differs from a humanoid's kinematics; see the human-to-humanoid embodiment gap.

Dataset-specific limits also apply:

• Benchmark overfit. A model tuned on a small catalog of known products may not transfer to unfamiliar designs.
• Cooperative setting. Participants who know they are recorded may follow the manual more closely than real assemblers do.
• Force blindness. Tightening and tapping are defined by force, which no camera records.
• Scale mismatch. Large panels and bimanual carrying stress a robot's payload and balance in ways small-part videos never reveal.

Permissions and provenance

Product designs, brand marks and manual illustrations may belong to the furniture maker or a licensor, and an assembly service's footage may show customers' homes and family members. Redaction or removing identifiers does not settle permissions, and blurring a face does not clear rights in the product or the manual. The U.S. Copyright Office has published a report on copyrighted works used in developing generative AI systems (U.S. Copyright Office, 2025); it is a starting point for the questions below, not an answer to them. This is not legal advice.

Questions to ask suppliers

1. Which products, materials and hardware types are covered, and are manuals available for each?
2. Who owns the manual and the product design, and can that party license them for training?
3. Were assemblers employees, contractors or customers, and what notice did they receive?
4. Do any recordings show customer homes or households?
5. Are mistakes, backtracking and tool changes preserved?
6. How was the video aligned to manual steps, and by whom?
7. Which camera views and mounts exist, and is any depth or tracking data available?
8. Are public benchmark products excluded so you can keep evaluation data separate?

Discuss your requirements

If furniture or flat-pack assembly is part of your evaluation or pretraining plan, describe the products, modalities and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match. See the humanoid training data guide and the manufacturing hub for neighboring tasks, and the catalog page on first-person video of skilled manual work for the related collection program.

Describe the workplace demonstration data your robotics project needs

Sources

• The IKEA ASM Dataset: Understanding People Assembling Furniture through Actions, Objects and Pose — video, action and pose annotations for furniture assembly.
• IKEA-Manual: Seeing Shape Assembly Step by Step — manuals paired with objects and assembly plans.
• Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — report on copyrighted works in generative AI development.