Short answer
Repair video shows technicians diagnosing and fixing equipment with hands, tools and parts in view, the visual counterpart to manuals and work orders. It can support tool and part recognition, step recognition and visual pretraining. It cannot show applied force or work hidden inside equipment, and filming on customer premises complicates rights. Existing footage is thin and uneven, so many buyers plan commissioned recording of selected repairs.
Why repair is harder to record than other tasks
Repair is long, staged and tool-heavy. A typical job moves through diagnosing, isolating, removing, replacing, testing and restoring, and the camera that sees the technician's hands often cannot see the fastener. Work happens in cabinets, under dashboards and behind panels, where headlamps blow out the frame and the technician stops to read a meter, a screen or a manual page. Compared with tabletop picking, much of the information sits where a camera cannot go.
That makes the supporting records important. A repair clip becomes far more useful when the work order, the procedure and the part list that surround it are linked to it. Work order histories hold the symptom, diagnosis and outcome; this page is about the footage.
What published skilled-activity video shows
- Ego-Exo4D recorded skilled activities, bike repair among them, simultaneously from an egocentric camera and several exocentric ones. The version cited here (arXiv v4) has 740 participants in 13 cities, 123 scene contexts and 1,286 hours, plus eye gaze, IMU, 3D point clouds, camera poses and multichannel audio, with expert commentary and benchmarks that include proficiency estimation (Grauman et al., 2023). Repair is one activity among several, not the whole corpus.
- Ego4D gathered 3,670 hours of daily-life video from 931 camera wearers in 74 locations across 9 countries, from consenting participants with de-identification procedures, and reports more than 250,000 hours of annotator effort (Grauman et al., 2021). The annotation effort dwarfs the recording time, which is a useful warning when you budget for labels.
- IndustReal includes assembly and maintenance procedures in egocentric video, with procedural and execution errors (Schoonbeek et al., 2023).
These are consented research recordings. They show what to ask a supplier for: synchronized views, expert speech, step structure and kept errors. They say little about how much usable repair footage businesses hold.
Where existing repair footage may come from
| Source | What it may hold | Strength | Watch for |
|---|---|---|---|
| Field-service body or helmet cameras | First-person footage of service visits | Real equipment under real conditions | Customer premises; continuous capture includes incidental private moments; see technician body-camera footage |
| Remote-assistance sessions | Live video with an expert's guidance | Instruction tied to steps | Framing follows the call, not the task; see remote-assistance session recordings |
| Manufacturer and dealer service training | Filmed procedures on specific models | Clear steps and lighting | Staged and edited; equipment-maker ownership; see OEM service documentation access |
| Repair shops and depots | Bench documentation, before-and-after clips | Many jobs of the same type | Shop and customer property; staff consent |
| Public how-to repair videos | Wide range of repairs | Breadth | Rights sit with creators and are subject to platform terms; see public instructional video vs licensed footage |
Shops and service networks as a supplier class are covered in repair shop networks as suppliers, and the catalog entry for first-person video of skilled manual work describes the related asset type.
Assessing a repair clip
Sample clips yourself and score them against these questions.
- Working steps in view. Are hands and tool tips visible while the work happens, or only during setup and cleanup?
- Continuity. Does the recording run through hard steps, or cut away at them?
- Interior visibility. Can you see inside the cabinet, and can you read part numbers, meter faces and screens?
- Diagnosis, not just replacement. Does the clip include tests and measurements, or only the swap?
- Narration. Does the technician explain what they are doing, and does tool noise mask it?
- Camera. Head, chest or shoulder mounted, with what field of view and stabilization?
- Known fault and outcome. Is the fault documented, and is the final test result recorded?
- Linkage. Can the clip be tied to a work order, an asset and a procedure revision? See SOP-to-video step alignment.
Commissioned recording of selected repairs
When existing footage is thin, record a small set of repairs chosen for frequency and duration. Work orders show which jobs recur; prioritizing maintenance tasks for collection covers the selection. Useful design choices include staging on real or retired equipment in a shop, seeding faults by installing a known failed part instead of damaging equipment, recording several technicians on several units for variation, adding a fixed bench or wide camera beside the head-mounted one, and asking for think-aloud narration. Run a pilot first; pilot collection scoping explains what a first run should prove. If torque or force matters, add instrumented tools. If no existing asset fits, SourceX can scope a commissioned collection of new recordings.
Illustrative package
# Illustrative only — synthetic equipment, identifiers and values; not an available dataset
package: repair_video_pack_v0
scope: 3 repair types on rooftop HVAC units (capacitor swap, contactor swap, condensate drain clearing)
recording:
views: [ego_head, chest, fixed_wide]
video: constant frame rate, original masters retained
audio: technician think-aloud; tool noise flagged
per_clip_metadata: [site_id, technician_id (pseudonymous), equipment_class, procedure_id + revision, work_order_id, fault_known]
annotations:
stage: diagnose | isolate | remove | replace | test | restore
step_segments: start and end time per procedure step
tools_and_parts: per-segment lists from a controlled vocabulary
visibility: per-second flag for hands and tool tip (visible | occluded | out_of_frame)
outcome: fixed | not_fixed | partial, with the final test result
excluded: applied force or torque, interiors blocked by panels, customer documents on screens
rights: training and internal evaluation; no redistribution; site-owner and technician consent on file
What repair video can and cannot support
Repair video can support tool and part recognition, stage and step recognition, grounding of written procedures, video-language pretraining, outcome estimation and evaluation of step prediction; see troubleshooting evaluation sets. It cannot supply robot joint states or actions, applied torque or force, or the steps hidden inside equipment. Whether a lockout was applied may not be visible at all. Diagnostic reasoning appears only where the technician narrates or the clip is paired with records. Published work uses human video alongside robot data rather than in place of it, as summarized in what human demonstration video can and cannot do. Tool-specific skill and grip are covered in hand tool use demonstrations, and the force gap in when video is not enough.
Permissions and provenance
Repair footage usually involves several parties. The customer or site owner controls the premises and what is in frame, including nameplates, addresses and documents. The technician is an employee or contractor whose consent should cover the specific purpose, not just a workplace notice. Equipment makers' manuals, diagrams and software screens may appear on camera, and footage showing unsafe practice raises its own questions. Blurring or masking does not resolve these permission questions. See worker consent and identifiable people, third-party content in recordings and confidential business information in footage. This is not legal advice.
Questions to ask suppliers
- Who recorded the footage, for what purpose, and under what notice to technicians and customers?
- What camera types and mounts were used, and can you see inside the equipment during the work?
- Do clips run continuously through the job, or are hard steps cut out?
- Can clips be tied to work orders, asset IDs and procedure revisions?
- Are failed repairs, callbacks and corrections kept?
- Do customer contracts allow recordings made on their premises to be licensed for AI training?
- Is any manufacturer-owned material, such as manuals or software screens, visible in the footage?
- Is there narration, and in which language and quality?
Discuss your requirements
If you need repair footage for specific equipment classes, describe the repairs, views and permitted uses to SourceX. SourceX looks for US businesses that hold suitable footage, checks the data and the supplier's licensing permissions, and coordinates pricing, allowed uses and delivery in a license. 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. SourceX reviews each request and follows up about potential sources; a request does not guarantee a match.
See also the maintenance, repair and tool-use hub, the humanoid training data guide, existing footage vs commissioned recording and the catalog entry for field service and maintenance work orders.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — bike repair among skilled activities, multi-view capture and expert commentary.
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — consent, de-identification and annotation effort in large egocentric collection.
- IndustReal: A Dataset for Procedure Step Recognition Handling Execution Errors in Egocentric Videos in an Industrial-Like Setting — maintenance procedures with errors in egocentric video.