Disassembly and Reassembly Sequences

The buyer's problem

Taking something apart is a planning problem before it is a manipulation problem. The order matters: a cover must come off before a connector is reachable, a battery must be disconnected before a board is lifted, and a clip that is broken on the way in cannot be reused on the way out. Reassembly then has to restore the same state, often with parts that have been cleaned, swapped or replaced. A humanoid that does repair, refurbishment or dismantling needs examples of both directions, including what goes wrong.

The sourcing question is where such sequences exist at scale in a form a model can use. Repair centers, refurbishers and remanufacturers do this work all day, but their records are often work orders and photos, not labeled sequences. This page covers what to look for, and how it connects to the broader repair video demonstrations and e-waste dismantling topics.

What published datasets show

Two public resources illustrate the data shapes involved, though neither is a repair-shop dataset.

Assembly101 recorded 4,321 videos of people assembling and disassembling 101 toy vehicles, using 8 static and 4 egocentric cameras, with more than 100,000 coarse and 1 million fine-grained action segments and 18 million 3D hand poses (Sener et al., 2022). The recordings were unscripted and include mistakes, and the paper introduces a mistake-detection task. The useful lesson for buyers is the combination: several viewpoints, hand tracking, fine segmentation and naturally occurring errors in one package.

IKEA-Manual pairs 102 IKEA objects with their assembly manuals and annotates decomposed parts, assembly plans, manual segmentation and 2D-3D correspondence (Wang et al., 2022). It shows how document-side structure, rather than video, can express a sequence. It covers assembly, so any reverse sequence is inferred, not observed.

Neither study tells you how many examples your model needs. That depends on architecture, product variety and what the model is meant to output.

What the data can and cannot support

Supplier types and what each holds

• Independent repair centers may see a broad range of real failures, but may not record video. New recording is often needed.
• Refurbishers and remanufacturers repeat the same teardown across many units of one model, which gives consistent sequences and a natural source of variation in wear and damage.
• E-waste and recycling operators disassemble many product types quickly, with less care for reversibility.
• Manufacturers hold design-side sequences and service manuals, covered under OEM service documentation access.

Existing footage, where it exists, is often a side effect of training or quality review. Commissioned recording lets you set camera count, step labels and tool lists, but costs more and needs the supplier's site access and worker consent. The tradeoffs are laid out in choosing a sourcing route.

Illustrative package for a teardown-and-rebuild recording

# Illustrative only - fields and values are synthetic examples, not an available dataset
package: teardown_rebuild_v0
product_family: small_appliance_class_A (12 unit models, 6 units each)
video:
  views: 2 fixed (top, front) + 1 head-mounted
  fps: 30 constant, 1080p
  sync: shared timecode, offset logged
annotations:
  step_segments: [step_id, verb, part_id, tool_id, t_start, t_end]
  state_snapshots: photo of part layout after each major step
  reversal_link: disassembly step_id mapped to its reassembly step_id
  events: [dropped_part, wrong_part, rework, broken_clip]
catalog_link: part_id to the supplier's parts list
excluded: robot actions, joint states, torque and force values
rights: training and evaluation use, worker consent on file, no redistribution

Limitations and additional preparation

Teardown video shows a human body doing the work. Hand shape, strength and reach differ from a robot hand, and the embodiment gap applies fully. Fastener resistance, adhesive release and clip deflection are invisible in video, yet they decide whether a step succeeds.

Sequences can also be product-specific. A model trained on one appliance family may learn that family's screw layout, not general teardown skill. Ask for the number of distinct models and variants, and keep a held-out split by product model, not only by unit.

Typical preparation includes aligning cameras by timecode, assigning stable part IDs, labeling step boundaries consistently across annotators, and marking the cases where the operator deviated from the service manual. If you intend to use reversibility as a training signal, verify it on a sample of real teardowns.

Permissions and provenance

Repair shops handle customer-owned devices, and a device on the bench may hold personal data, serial numbers or a customer's name on a tag. Ask whether footage shows screens, labels or documents, whether customers consented to recording, and whether the product's design is the manufacturer's confidential material. Blurring a screen does not settle whether the shop may license the footage. This is not legal advice.

Quality and acceptance criteria

Write acceptance tests into the order so that disputes have a reference. A reasonable set, with thresholds that you set from your own use, covers six areas.

• Step coverage. Every unit has a complete disassembly and a complete reassembly, with no unexplained gaps between segments.
• Label agreement. A second annotator relabels a sample of units, and agreement on step boundaries and verbs is reported.
• Part traceability. Each removed part has an ID that persists across the two directions, so a screw taken out at step 4 can be matched to its reinsertion.
• Viewpoint coverage. Hands and the working area are visible for a stated share of each clip, measured on a random sample.
• State records. Photos or states after major steps allow you to check that reassembly returned the product to its starting configuration.
• Error annotation. Dropped parts, wrong orders and rework are tagged, since a set with only clean runs teaches little about recovery.

If any criterion fails on the sample, return it before the full delivery, not after.

Sequences can also be cross-checked against drawings and plant records: exploded views and parts catalogs give part structure but not order, and assembly sequence data covers the production-line counterpart. Tool handling within each step is covered in hand tool use demonstrations, and finding businesses that do this work in repair shop networks as suppliers.

Questions to ask suppliers

1. How many distinct product models and variants are covered, and how many units per model?
2. Is the teardown sequence the shop's standard procedure, a manufacturer procedure, or ad hoc?
3. Are both disassembly and reassembly recorded for the same units?
4. Do recordings include failed steps, broken parts and rework, and are they labeled?
5. How many cameras were used, and can the shop provide camera placement and timecode details?
6. Can steps be linked to part numbers or a parts catalog?
7. Were customer devices recorded, and what do customers sign?
8. Who owns the product design information visible in the footage?
9. Can the shop record new sequences for a specified product family, and what site access does that need?

Discuss your requirements

If you are scoping teardown and rebuild data for a product family, describe the products, camera setup 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. For context, start at the maintenance data hub or the humanoid training data guide, and see the catalog entry for first-person video of manual work.

Describe the workplace demonstration data your robotics project needs

Sources

• Assembly101: A Large-Scale Multi-View Video Dataset for Understanding Procedural Activities — multi-view assembly and disassembly video with hand poses and mistakes.
• IKEA-Manual: Seeing Shape Assembly Step by Step — manual-paired assembly plans and part annotations.