Short answer
Interior finishing demonstrations record tasks such as hanging and fastening drywall, taping, sanding, painting and installing fixtures. They can support task and step understanding, spatial reasoning about large panels and long-horizon evaluation. They do not provide robot actions, tool forces or site-specific plans, and every building differs, so coverage across sites matters more than hours from one.
Why interior finishing is distinct
Interior finishing is large-format, tool-heavy and variable. A drywall sheet is large, awkward and fragile at its edges. Painting depends on surface preparation, coverage and drying. Fixtures such as doors, cabinets and lights must align to openings, studs and utilities that differ by building.
Compared with factory work, there is no fixed station. The workspace is a partly built room with dust, cables, scaffolds and other trades. The recording challenge is also different: sites restrict access, schedules slip, and safety rules govern what can be worn or carried. A buyer asking for "construction robot training data" should narrow to a trade and a task, because the data needs of drywall hanging, painting and fixture mounting have little in common.
Task anatomy
| Task | Typical steps | What makes it hard |
|---|---|---|
| Drywall hanging | Measure, cut, lift, position, fasten at studs | Panel size and weight; two-person lifts; alignment to framing |
| Taping and finishing | Apply tape, mud, feather edges, sand | Surface judgment; dust; repeated coats |
| Painting | Mask, cut in edges, roll, inspect, touch up | Coverage, drips, lighting changes |
| Fixture installation | Locate, drill, anchor, mount, level, connect | Accurate layout; tool changes |
| Cleanup and handover | Remove masking, collect debris, punch-list fixes | Dispersed small tasks |
Failure modes: panels cracking at edges, screws set too deep, tape bubbles, roller marks, drips on masked areas, fixtures out of level, and rework after an inspector's punch list.
Using a task taxonomy
O*NET includes task statements and detailed work activities for occupations, and a file that maps tasks to detailed work activities; task statements are retained only if the relevance rating is 25% or higher (O*NET Resource Center). A buyer can use it to name tasks consistently when comparing suppliers. It does not describe motion or tools at robot-relevant detail, so it supports vocabulary and coverage planning, not step segmentation. For a robot-oriented scheme, build an action taxonomy for workplace video on top of it.
What the data can support
- Task and step recognition. Video with step labels trains models to follow multi-stage trade work.
- Spatial and layout understanding. Where studs, openings and fixtures sit relative to a panel or tool.
- Long-horizon evaluation. Finishing tasks run for hours with state that persists, such as coats applied; see long-horizon task demonstrations.
- Two-person and carrying behavior. Lifts and handoffs show coordination; see manual load handling and carrying.
- Method statements as language. Written construction method statements can be paired with video to give step-level language, though a method statement describes intent rather than what was done.
- Evaluation of quality outcomes. Before-and-after images with inspection results help define success labels.
What it cannot support
- Robot actions and tool forces. Pressing a drill, loading a roller, feathering joint compound and pulling a panel flush involve forces not visible in video. See when video is not enough.
- Embodiment match. A person lifting a sheet with a shifting grip and body lean cannot be replayed by a humanoid; see the embodiment gap.
- Metric site geometry. Hand-held video lacks a floor plan and calibrated depth unless captured. Without it, "place the panel here" cannot be tied to coordinates.
- Mobile base and navigation. Moving through a half-built room is a separate problem; see mobile manipulation demonstrations.
- Coverage of the happy path only. Active sites produce clean successes in the footage that gets approved for release.
Capture on live sites
Recording during real work needs a plan agreed with the general contractor and the trade foreman. Cameras on hard hats are common, but each site has policies on photography, and other trades may be in view. Recording protocol on live work sites covers the practical items. Depth and floor-plan capture, such as a room scan before work begins, make the video much more useful for spatial learning. Consider masonry and tiling for sibling trades with different materials.
Illustrative collection spec
# Illustrative only; synthetic values
dataset: interior_finish_v0
trades: [drywall_hang, tape_and_finish, paint_roll_and_cut, fixture_mount]
sites: 6 (residential, commercial tenant fit-out), 4 contractors
capture:
head_camera: hard-hat mount, constant fps, wide field of view
fixed_camera: tripod, room-scale, optional
site_scan: pre-work room scan with dimensions
labels:
task_level: trade, phase, step segments
objects: panel_id, fixture_id, tool type
outcome: inspected_ok / rework / punch_list_item
conditions: lighting, dust level, other_trades_present
splits: by site and contractor
excluded: robot actions, tool forces, structural drawings
Acceptance criteria
- Each recording ties to a site ID, trade and task, with phase boundaries labeled.
- Camera placement and field of view are documented and consistent within each site.
- Outcomes come from an inspection record, not from the annotator's impression.
- Hard cases (cutouts, corners, ceilings, awkward lifts) are present, with counts.
- Evaluation sets are held out by site and contractor.
Permissions and provenance
Construction sites involve several parties: the property owner, the general contractor, specialty subcontractors and individual workers. Ask who can authorize recording and licensing; a subcontractor's footage on someone else's project may be subject to the project contract. Footage may show other trades, visitors, client belongings or drawings. Worker identifiability is a concern; see worker consent and identifiable people. The EDPB's guidelines on video devices are one reference on notice and proportionality (EDPB, 2020). This is not legal advice, and blurring does not settle who may license the content.
Questions to ask suppliers
- Which trades and tasks are recorded, and over how many sites?
- Who is the contract party for each site and does it allow recording for AI training?
- Are there site scans, dimensions or drawings that can be licensed alongside the video?
- Do recordings include rework and punch-list corrections?
- What camera mount and field of view were used?
- Can the supplier link video to project records such as construction project records?
- What safety rules governed the recording, and were they followed?
- Are workers and visitors identifiable, and how was consent handled?
Discuss your requirements
If you are scoping interior-finishing data, describe the trade, tasks, modalities and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If nothing existing fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match; SourceX follows up about potential sources.
Start with the humanoid training data guide or the workplace task demonstration hub. Related catalog entry: first-person video of skilled manual work.
Describe the workplace demonstration data your robotics project needs
Sources
- ONET Database (ONET 29.1 per search snapshot): Task Statements and Detailed Work Activities — O*NET task statements and detailed work activities as a task taxonomy
- Guidelines 3/2019 on processing of personal data through video devices (Version 2.0) — Guidance on personal data processed through video devices