Imagine a skilled worker who spent years learning how to adapt movements to small material differences. They are paid to demonstrate for robot training. Years later, similar capabilities earn revenue across many locations.
That does not automatically give them a legal right to income from every machine task. It raises another question: what exactly did the demonstration agreement price?
A few hours of labor, a license to use data, or a capability that can be copied repeatedly?
Physical AI gives experience a new means of transmission. It also calls for fresh thought about how that contribution is traded.
Physical Intelligence's account of π0.5 shows the role of multiple knowledge sources and training data in robot generalization, while acknowledging errors in understanding and action. Data dependence does not establish that one person owns a movement, nor a general right to subsequent royalties.Research team's explanation
What is being traded in a demonstration?
Before discussing returns, separate the relationships involved.
Demonstration pay compensates time, skills, and working conditions. Completed work retains its value even if the model never succeeds commercially.
A data-use license specifies permitted purposes, users, transfers, and duration. Actual rights also depend on contracts, employment relationships, content, and applicable law. Performing a movement does not automatically confer complete ownership of all related data.
Skill-service fees can pay for ongoing annotation, quality improvement, and adaptation to new settings. That is continuing delivery, rather than charging indefinitely for one demonstration.
A share of future returns is a separately agreed arrangement: contributors may receive bonuses, revenue shares, or asset returns through suitable structures when conditions are met. It offers potential upside alongside delays, uncertainty, and enforcement costs.
These arrangements can be combined or chosen after proper explanation. Vague future participation should not offset clearly promised current labor pay. A promising-sounding future share must not make work already performed nearly free.
How much value does one movement contribute?
An immediate objection is that robot capabilities come from many demonstrations, architecture, compute, engineering, safety testing, customer deployment, and capital. Why single out data contributors for distributions? How can anyone attribute revenue to one movement?
That question deserves a serious answer.
Models combine many sources. A successful task rarely traces to one unique demonstration. A sample's use in training does not establish its separately measurable marginal value. Upload counts are easy; attributing returns is hard. On-chain records can improve consistency without solving attribution.
Start with arrangements that can be enforced
I do not think the next step is a perpetual royalty for every movement. Simple contracts around identifiable collaborations are a more practical beginning.
One option is adequate one-time pay. For a defined, time-limited collection project, a larger guaranteed payment may best meet contributors' needs. Choosing certainty is not shortsighted: many households cannot afford to bet on whether a startup succeeds.
Another is base pay plus milestone bonuses. Participants receive additional payments when agreed data-quality, coverage, or acceptance targets are met. Milestones must be independently checkable, rather than changeable at the payer's discretion.
A third is a collective pool. Where individual attribution is difficult but the collaborating group is clear, an agreed share of defined revenue could fund rewards distributed under rules agreed in advance. This avoids false precision, while still requiring decisions about membership, entry, exit, and internal fairness.
A fourth is continuing collaboration. Involving skilled practitioners in error analysis, testing, and new settings may preserve bargaining power better than one-off collection. They would be paid for continuing to provide valuable judgment, alongside any return on their earlier data.
These are negotiable options, not proven best practices for every industry. Before implementation, examine whether administrative costs consume the benefit and whether contributors can understand and enforce the agreement.
Long-term revenue sharing needs clear accounts
A long-term sharing agreement should specify the product or business covered, revenue or profit basis, deductible costs, payment frequency, verification, and treatment of renaming or business transfers.
Profit sharing can sound generous but mean little if costs are allocated arbitrarily. Pay per use can sound exact but be unworkable if the system cannot identify the relevant use. Sound mechanisms begin with auditable accounts.
Be precise about permission and withdrawal
Permission must also consider other people and organizations. Demonstrations may contain customer information, household footage, or employer trade secrets; one participant's consent does not authorize every use. Withdrawal also needs precision: stopping collection, deleting raw data, and removing influence from a trained model are technically different. Do not promise one-click erasure of all three.
Give contributors meaningful choices
Addressing these questions is more than a cost for companies. Clear scope, dependable compensation, and feedback channels may attract skilled contributors and sustain partnerships. That effect requires project-specific validation; fairness is not a guarantee of higher profits.
Start with a manageable pilot: guarantee base pay, offer informed choices among reward structures, and record understanding, disputes, administrative costs, and continued collaboration. Publish aggregate results at the end before expanding. Do not issue a token first and decide its rights later.
The previous essay examined labor behind machines; this one examines the reproducible capabilities labor leaves behind. Together they ask whether we will clarify contribution, compensation, and risk before technology spreads.
Experience can help machines develop. Institutions and contracts can give the people providing that experience more choices.
If your skills were used to train robots, would you prefer higher one-time pay or base pay plus long-term participation? What information would change your choice?
Sources current to October 20, 2026. This essay proposes collaboration mechanisms; it does not assert universal automatic property rights or statutory royalty entitlements.