Short answer
Maintenance and repair training data for robots combines records of what failed and how it was fixed, documentation of how service should be done, and video of technicians using tools. It supports diagnosis, repair-step prediction, tool and part recognition and procedure grounding. It does not capture how much torque a fastener took or how a tool felt in the hand, so tool control needs instrumented or robot data.
Scope of this cluster
Maintenance is where a humanoid would combine reasoning and dexterity most tightly: decide what is wrong, choose a procedure, pick the right tool, then use it in a cramped, greasy, poorly lit space. This cluster covers the data for each part. It owns tool use generally, so generic tool-handling pages live here rather than in workplace task demonstrations.
Three data families recur:
- Records. Work orders, CMMS exports, maintenance logbooks, preventive-maintenance checklists and callback history. They link a symptom to a diagnosis, parts and outcome.
- Documentation. Service manuals, troubleshooting guides and fault trees, torque specifications, exploded views, wiring diagrams and lockout/tagout procedures.
- Demonstrations. Repair video, technician body-camera and remote-assistance sessions, and recordings of hand and power tool use.
They sit in different layers of the humanoid training data guide: records and documents feed reasoning and planning, video feeds perception and task understanding, and none of them feeds low-level tool control on its own.
What each family can support, and what it cannot
| Data family | Supports | Cannot support alone |
|---|---|---|
| Work orders and logs | Diagnosis and outcome reasoning, task frequency for prioritizing, maintenance language | Any visual or physical signal |
| Service documentation | Step structure, parameters such as torque values, tool lists, safety steps | Execution: motion, force, feel; licensing is often blocked by the equipment maker |
| Repair and tool-use video | Tool and part recognition, step recognition, visual priors for tool use | Applied torque, contact force, what happened inside an occluded space |
| Instrumented tools and rigs | Force and torque during human tool use | Workplace scale; it is specialist capture |
Records are messy text. MaintNet, a library of maintenance language resources, ships logbook data from aviation, automotive and facilities alongside abbreviation lists and termbanks, which signals how much normalization technician shorthand needs before a model can use it (Akhbardeh et al., 2020). Budget for that work when licensing CMMS exports.
Video can show skill as well as steps. Ego-Exo4D recorded skilled activities, including bike repair, from egocentric and exocentric views with expert commentary, and includes proficiency estimation among its benchmarks (Grauman et al., 2023). IndustReal covers assembly and maintenance procedures and includes procedural and execution errors (Schoonbeek et al., 2023). Both are research corpora, but they show what to ask of business footage: a stable view of the hands, step structure, and errors that are kept rather than edited out.
Force is the hard gap: video does not record whether a bolt reached specification or how a seized fastener resisted. Portable hand capture does not solve this either: DexCap tracks hand motion but records no force (Wang et al., 2024). Contact-rich robot datasets such as RH20T log force with every sequence (Fang et al., 2023), and instrumented tools can capture force and torque during human work. For when this matters, see when video is not enough.
Sourcing routes
Existing records. Field-service organizations, facilities teams, fleets and MRO providers hold years of work orders. Customer names, site addresses and technician identities need redaction, and asset IDs need to survive it so records can be joined. Start with work order histories; the catalog entry for field service and maintenance work orders covers the record type.
Documentation. Service companies hold manuals but rarely own them. Equipment makers and their dealer agreements usually control service documentation, so a repair business often cannot license it; see OEM service documentation access.
In-house procedures are a different case. US employers must develop, document and use energy-control procedures when employees service equipment, with a narrow documentation exception, per OSHA's guidance on 29 CFR 1910.147 (OSHA), so lockout/tagout procedures are a common business-owned document. This is not legal advice.
Video. Field-service firms that use body cameras or remote assistance may keep recordings, but these are filmed on customers' premises, which adds a permissions layer. Repair shops and dealer service departments are another route. Where footage is thin, commissioned recording of tool use can be scoped; if no existing asset fits, SourceX can scope a commissioned collection.
Specialist capture. Force and torque during tool use come from instrumented tools 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.
How to evaluate maintenance supply
- Join keys. Can work orders be linked to assets, procedures and, where it exists, video?
- Closure quality. Share of work orders with a coded cause and fix rather than "done" or "see notes."
- Outcome labels. Whether callbacks and repeat visits are recorded, so first-time-fix can be computed.
- Document ownership. For each manual or guide, who wrote it and who may license it.
- Framing. In repair video, how often the tool tip and fastener are visible in tight spaces.
- Premises permissions. Whether customer contracts allow recordings made on site to be used for AI training.
- Safety content. Whether lockout steps and hazards are present and current.
Illustrative: one repair across three data families
| Source (synthetic) | What it contributes | What it lacks |
|---|---|---|
Work order WO_2026_04417, asset RTU_118 | Symptom "no cooling," diagnosis "failed capacitor," part number, labor time, callback flag | Any image of the unit or the work |
Procedure PM_HVAC_022 rev D | Lockout steps, access panel removal, capacitor test and replacement, tool list | How the technician held the meter leads; panel screw torque |
Body-camera clip bc_tech_A07_0311 | Hand and tool views for 11 of 14 procedure steps | Steps hidden inside the cabinet; applied force; consent from the site owner still to confirm |
Joined, the three give a planner the diagnosis and procedure and give a perception model labeled tool-use footage. A control policy for the same repair still needs robot or instrumented data.
Key pages in this cluster
| Group | Page | What it answers |
|---|---|---|
| Scoping | Equipment-specific vs transferable skills | Which maintenance skills generalize |
| Scoping | Prioritizing maintenance tasks for collection | Choosing tasks worth recording |
| Records | Work order histories | Symptom, diagnosis, parts and fix |
| Records | Maintenance logbook text | Free-text logs and their noise |
| Records | Preventive maintenance checklists | Routine tasks and completion records |
| Records | CMMS asset hierarchies | Equipment registers that make records joinable |
| Records | First-time-fix and callback labels | Outcome labels from repeat visits |
| Records | Troubleshooting guides and fault trees | Structured diagnostic logic |
| Documentation | Service manuals for embodied AI | What manuals give a humanoid |
| Documentation | OEM service documentation access | Who can license equipment-maker documents |
| Documentation | Torque specs and fastening procedures | Parameters for fastening tasks |
| Documentation | Exploded views and parts catalogs | Part structure and identity |
| Documentation | Lockout/tagout procedures | Safety procedures as data |
| Video and tool use | Repair video demonstrations | Sourcing repair footage |
| Video and tool use | Remote-assistance session recordings | Guided repairs with expert talk |
| Video and tool use | Technician body-camera footage | Evaluating first-person field footage |
| Video and tool use | Hand tool use demonstrations | Wrenches, screwdrivers and pliers |
| Video and tool use | Power tool use demonstrations | Drills, drivers and grinders |
| Video and tool use | Tool identification image data | Recognizing tools |
| Video and tool use | Measuring instrument use | Calipers, multimeters and gauges |
| Video and tool use | Tool affordances and grasp annotations | Where and how tools are held |
| Video and tool use | Instrumented tools | Capturing force and torque during tool use |
| Evaluation | Troubleshooting evaluation sets | Repair-step prediction tests |
| Evaluation | Diagnostic reasoning traces | How technicians reason to a fix |
Neighboring clusters
Written procedures in general are covered by the SOPs and procedural data hub. Machine tending and changeovers sit in manufacturing and assembly; equipment wear and damage images in inspection and failure data; spare-part recognition overlaps product and object data. For the record type in the catalog, see licensing maintenance logs.
Discuss your requirements
Tell SourceX which equipment classes and tasks you target, which of the three data families you need, and whether you need force data. Describe the maintenance data to SourceX: it investigates whether US businesses hold suitable records or footage, checks 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
- MaintNet: A Collaborative Open-Source Library for Predictive Maintenance Language Resources — maintenance logbook data and normalization resources.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — repair among skilled activities, with commentary and proficiency.
- IndustReal: A Dataset for Procedure Step Recognition Handling Execution Errors in Egocentric Videos — maintenance procedures with errors.
- 29 CFR 1910.147 The control of hazardous energy (lockout/tagout) — documented energy-control procedures.
- DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — hand capture without force sensing.
- RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot — force recorded with contact-rich robot sequences.