Ordinary video gives incomplete scale and depth unless it is paired with measurements or depth capture.
Research direction
Video is evidence—not a quote and not a robot command.
The research question is narrower and more useful: how much of a physical job can be converted into a verified work record before the work, and how quickly can that record learn from what actually happened?
Near-term system
A staged path from observation to a reviewed estimate.
- 1. CaptureGuided site video, spoken context, photos, and known measurements.
- 2. ObserveSeparate visible evidence from assumptions and missing information.
- 3. StructureDraft work areas, tasks, dependencies, hazards, tools, materials, and stop conditions.
- 4. ConfirmA competent human corrects the record and owns safety, scope, and price decisions.
- 5. CompareRecord actual labour, materials, surprises, cost, and outcome after the job.
The durable asset
A before-and-after dataset grounded in completed work.
Each consented job links the initial instruction and site evidence to the confirmed task plan, actual sequence, people, tools, materials, duration, cost, surprises, and before/after outcome.
Predictions are stored with confidence and evidence. Corrections are part of the data, not hidden as product failure.
Why video alone is insufficient
The camera cannot see through walls or feel the work.
Services, substrates, moisture, fixings, and prior work may be invisible until inspection or disassembly.
Force, resistance, sound, smell, texture, and tool response matter during execution.
Safety rules, building context, access, customer expectations, and trade practice still need a responsible person.
Longer-term robotics path
Confirmed task graphs before supervised machine skills.
A research robot would receive a bounded task, world-state evidence, allowed tools, constraints, verification steps, and stop conditions. It would not receive a customer video and improvise the whole job.
The credible sequence is dataset → verified task graph → simulation → one bounded skill → supervised execution. Pricing research and robotics research can share the work record without pretending they are the same product.
Research safeguards
Consent, minimisation, and human responsibility by default.
Only authorised job data is accepted. Participants can avoid faces, personal papers, number plates, and unrelated rooms. Accepted recordings use a private transfer path, limited access, and a stated retention purpose.