Worked example: pricing a rare manipulation dataset
示例
Givens (hypothetical): a reseller holds a teleoperated dexterous-manipulation dataset. It contains 2,000 episodes of a rare task (tight-clearance insertion) collected across 12 sites. Labels are verified: each episode has action labels, task segmentation, and a success/failure outcome, with a documented QC pass. Provenance records exist for every capture, and the license and consent scope explicitly permit third-party resale for commercial training. The target buyer is a foundation-model team that needs rare manipulation skills and has its own labeling pipeline but no access to this task. Assume the reseller's own collection cost was substantial and that comparable rare-task datasets are scarce.
Question: which attributes drive the price, and which pricing model fits?
演示讲解
Step 1 — Identify the binding attribute. The task is rare and hard to reproduce, so uniqueness is the dominant driver. Scale is modest (2,000 episodes), but for a rare skill, coverage of the task matters more than raw hours.
Step 2 — Check the risk-reducing attributes. Labels are verified and QC-documented, so the buyer does not have to re-label. Rights are clear: license, consent, and provenance all permit commercial resale. Both attributes remove risk the buyer would otherwise absorb, which supports a higher price.
Step 3 — Match the buyer segment. The buyer is a foundation-model team that already has a labeling pipeline. They value the rare task coverage and the clean rights more than the labels themselves, but the verified labels still reduce their integration cost.
Step 4 — Choose the pricing model. Because the value is concentrated in a rare, finite asset with clear rights, a premium per-episode price with a tiered license fits: a base tier for internal research, a higher tier for commercial training, and an exclusivity option priced separately. A subscription would underprice a finite rare asset; an as-is one-time sale would ignore the rights tiers the buyer may want.
Step 5 — Sanity-check. If the rights were unclear, the same dataset would fall to a low as-is price or no sale, because the buyer could not legally train on it. If the labels were unverified, the price would drop by the buyer's re-labeling cost.