Idea 23/Robotics & Manufacturing/Economic + technical proposal/2 min

The Shift Is the Dataset

Humanoids are entering factories to work. The valuable output may be the trajectory log, not the parts.

Factory deployments should be structured as paid physical-data collection, with the operator retaining rights to manipulation trajectories, because the marginal value of real embodied data now exceeds the marginal value of the labour.

01What changed

Humanoid robots entered Chinese factory floors in 2026, with deployments shifting from demonstration videos to production-line roles.1 Xinhua, 26 October 2026
Vision-language-action models became spatially guided generalist policies, and sim-to-real frameworks matured, but every one of them reports real-world interaction data as the binding constraint.2 HyperSim, IEEE 2026

02The gap

Robot deployment contracts are written as labour contracts: the vendor is paid for hours worked. Nobody prices the data, which means nobody designs the deployment to produce good data, and nobody shares the upside with the factory that supplies the environment.

03The idea

A shift-as-dataset contract has three parts: the deployment is instrumented to a fixed standard, the operator retains a licence to the trajectory data, and the factory receives a lower hourly rate in exchange for the data right and for tolerating instrumentation.

The technical contribution is a data-quality standard for embodied manipulation: coverage of contact states, force ranges, object variation, failure distribution, and annotation depth. Without a standard, robot-hours are not comparable and a data market cannot form.

The measurable question is a scaling law. How much task performance does a thousand robot-hours buy, and does the exponent differ between tasks where contact is incidental and tasks where it is the whole problem? An exponent near zero for contact-rich tasks would say that data alone will not get us there and that hardware or simulation must improve.

The economic prediction is that the data right becomes worth more than the labour as models improve, which inverts the bargaining position of factories over a few years and should be written into contracts now rather than discovered later.

04Why now

Humanoids are on real lines, VLA models are data-starved, and no standard for embodied data exists. The window to define the contract is before deployments become routine.

05The decisive test

Do this

Instrument three deployments across two tasks, log to a proposed standard, and fit a scaling law relating robot-hours to task success at fixed model size.

Success looks like

A reproducible exponent above 0.1 for manipulation tasks, with contact-rich tasks showing a measurably lower exponent and a standard that two independent labs can apply consistently.

Kills it

If performance saturates after a few hundred hours, data is not the bottleneck and the whole economic argument collapses.

06What would break it

Factories will resist instrumentation for confidentiality reasons.
Data licensing law for embodied trajectories is unsettled.
A negative scaling result would be commercially unwelcome and hard to publish.

07If it works

A market for physical-world data forms with a quality standard, and factories capture value from the environments they already own.