Short answer
Difficult and failed repairs, such as a seized fastener, a rounded head or a part that breaks during removal, are rare in routine footage but central to evaluating a maintenance robot. Repair networks may hold records and some video of them; the rest must be commissioned. The material supports recovery-step understanding and evaluation design, not robot actions or forces.
Why failed repairs are a separate purchase
A product defect is something wrong with the thing a customer bought. A repair failure is something that goes wrong while a technician works on it: a bolt that will not turn, a thread that strips, a plastic clip that snaps, a replacement part that does not fit, a fix that comes back as a callback. The failure belongs to the interaction between tool, fastener, material and person, so defect image datasets are the wrong place to look for it.
For a humanoid maintenance program the distinction matters because the happy path is the cheap part of the data. Routine removal of a bolt that turns freely appears in almost any repair video archive. The moment a system must decide to stop, change tools, ask for help or escalate is far less often recorded, and it is where an evaluation set earns its keep.
The failure episodes worth distinguishing
The names below are working categories a buyer can adapt. They are not a published taxonomy, and which ones matter depends on the equipment family and the robot's task.
| Episode type | What the recording would show | Typical label question |
|---|---|---|
| Seized or corroded fastener | Attempt, resistance, switch of tool or method | At which step did the technician change approach, and to what? |
| Damaged fastener head | Tool slipping, head deforming | Did the damage happen before or after the first attempt? |
| Broken fastener or part | Fracture during removal or refit | Was the break anticipated, and was the part recoverable? |
| Stripped or cross-threaded thread | Misaligned start, loss of hold | Was alignment checked before turning? |
| Wrong or ill-fitting part | Mismatch discovered at refit | Where was the mismatch detected, and how late? |
| Repair that returns | Original job plus later callback record | What differed between the first visit and the fix that held? |
A single hour of footage may contain none of these. A work-order system, by contrast, may flag them indirectly: extra labor lines, a second visit, a parts return or a note that says the part was "seized". That is why this page treats the records and the video as two halves of one package. See first-time-fix and callback labels for how the outcome side is usually encoded.
What existing assets may hold
Repair shops, dealer service departments, equipment rebuilders, fleet maintenance teams and franchised repair networks often hold work orders with free-text technician notes. Some also hold training videos or incident photos taken when a job went wrong, and a few record bays on camera for liability reasons. Whether any of this is usable is a diligence question, not a given. Repair-shop networks as suppliers covers how to approach them, and work order histories covers the record structure.
The text side is where the cleanest preparation work exists. MaintNet collects logbook data from aviation, automotive and facilities domains, with abbreviation lists, termbanks and pre-processing tools (Akhbardeh et al., 2020). It is a language resource, not a repair-failure dataset, so treat it only as a reference for how to normalize the notes that describe a failed job. The page on normalizing technician shorthand picks up that step.
Expect three weaknesses in existing material:
- Survivorship. Jobs that fail badly may be quietly redone, discounted or never logged in detail.
- Missing video at the moment of failure. Notes may say a bolt sheared, while the video shows only the finished repair.
- Thin cause labels. "Seized" might mean corrosion, cross-threading or a wrong tool; the record rarely separates them.
What commissioned collection can add
A commissioned program can plan for difficult jobs instead of waiting for them, using the same step logic as disassembly and reassembly sequences but adding the branches where the sequence breaks. A shop might set aside equipment with aged fasteners, or a team might stage a difficult removal on a test fixture. Both are legitimate, and both must be disclosed in the metadata: a staged seized bolt is not a natural one, and a model trained or evaluated on staged examples should be able to tell the difference.
Commissioning also lets the buyer require the things existing footage lacks: a viewpoint on the fastener and tool tip, a clock shared between cameras and any instrumented tool, and a technician who narrates why they are changing method. Instrumented tools can add torque or force traces, which video alone cannot supply. What commissioning cannot do is manufacture rarity on a schedule; natural failures occur at their own rate.
Safety and method belong to the shop. Heat, penetrating agents, impact tools and extraction methods carry real risk, and this page does not describe or endorse any practice. Ask suppliers to follow their own documented procedures and to record any deviation as part of the label.
Illustrative record for one failure episode
# Illustrative only - synthetic values, not an available dataset
episode_id: fail_0042
equipment_family: family_C # buyer-defined grouping
site_id: site_017
operator_id: operator_A12 # pseudonymous
failure_type: seized_fastener
natural_or_staged: staged # required field
video:
views: [wrist_cam, bench_overhead, tool_tip_closeup]
sync: shared_clock_ms
steps:
- {t_start: 12.4, label: first_removal_attempt, outcome: no_movement}
- {t_start: 31.0, label: change_of_method, to: other_tool_or_approach}
- {t_start: 96.2, label: resolved_or_abandoned, outcome: abandoned}
linked_record:
work_order_id: wo_00871
callback_flag: false
free_text_note: normalized_and_original_both_kept
force_or_torque: absent # present only with instrumented tools
The point of the layout is to keep the failure, the decision to change approach and the outcome as separate fields, so an evaluation can ask whether a system recognized the failure and what it did next.
What this data cannot support
- Robot control. Training a humanoid to turn, pry or extract needs robot or teleoperation data, not footage, written records or recognition labels alone. Human video of a difficult removal has no joint states, commanded actions or contact forces.
- Force-sensitive judgment. Whether a fastener is about to shear is largely a matter of feel. Video shows the slip afterward, rarely the cause.
- Rate estimates. A collected set of failures says nothing about how often they occur in the field, because collection is selective.
- Generalization across equipment. Failures are specific to materials, corrosion, age and fastener type; see equipment-specific versus transferable skills.
Good uses are narrower: building held-out scenarios for troubleshooting evaluation sets, teaching step segmentation that includes retries, and checking whether a planner knows when to stop and escalate. Keep evaluation scenarios separate from anything used for training, split by site or equipment family as the claim requires.
Permissions and provenance
Failure recordings often show customer-owned equipment, vehicle identifiers, shop staff and sometimes liability-sensitive moments. Ask who owns the footage, whether staff were notified, and whether the shop can license content that depicts a customer's property. Blurring faces or plates does not by itself settle those questions. For the wider consent picture, see worker consent and identifiable people. This is not legal advice.
If any technician performance reviews exist alongside the footage, keep them out of scope. A package for evaluating robot behavior should not carry assessments of individual workers.
Questions to ask suppliers
- How are failed or difficult jobs recorded: free-text notes, labor-overrun flags, callback links or none?
- Does any video or photo capture the moment of failure, or only the before and after?
- For seized or broken fasteners, can the record separate corrosion, damaged heads, wrong tools and misalignment?
- Were any examples staged or reproduced on purpose, and is that marked?
- Can first-visit and callback records be linked for the same asset?
- Which equipment families, fastener types and sites are represented, and how many distinct technicians?
- Who owns recordings that show customer equipment, and what notice were staff given?
- What redaction was applied, and does it hide the fastener, tool tip or hands?
- Is any torque, force or tool telemetry recorded for the difficult steps?
Discuss your requirements
If your evaluation or training plan depends on repairs that went wrong, describe the equipment families, failure types and permitted uses to SourceX. SourceX reviews the requirement, investigates whether US businesses hold suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. For the wider map, see the humanoid training data guide and the maintenance data hub; related catalog context is in field service and maintenance records.
Tell SourceX what your evaluation set needs to cover
Sources
- MaintNet: A Collaborative Open-Source Library for Predictive Maintenance Language Resources — logbook text resources and pre-processing; a language reference only, not a repair-failure dataset.