Serve Robotics announced Physical AI Services on November 5, 2026. The US company describes Serve Data Curator as a tool for finding rare, valuable moments in robot-collected data for model improvement. That purpose makes the selection rule important: a collection built to surface unusual cases has a different denominator from the recordings it came from.
Label the collection’s purpose
A team may deliberately gather difficult clips to inspect a recurring problem or build a focused training collection. That can be useful even when the collection is unlike normal operations. The measurement question is separate: what does the fraction of matching clips in that selected collection actually describe?
It describes the collection under its selection rule. Without information about the source pool and how clips were chosen, it cannot establish how common the same situation was across the fleet. This is a general interpretation of selective sampling, not a claim about a reported Serve failure rate.
Work through a fictional clip pool
Suppose a fictional team fully labels 1,000 unique clips from a defined recording period and finds five matching a particular scenario. It then builds a ten-clip review set containing all five matches plus five other clips. The scenario appears in 50% of the review set: five divided by ten.
In the fully labeled source pool, the corresponding clip fraction is 0.5%: five divided by 1,000. Both percentages are correct, but they answer different questions. Reporting the review set’s 50% as the source pool’s frequency would replace the denominator after selection.
Neither percentage is automatically an event rate per operating hour. Clips can differ in duration, overlap or show the same incident more than once. This invented example assumes unique clips and complete labeling only to isolate the effect of selection; none of its counts describes Serve’s robots.
Keep a small record with the export
- Purpose: state whether the set is for targeted review, training or estimating a defined frequency.
- Source pool: retain the recording period, included machines and counting unit.
- Selection: preserve the query or rule and note any manual additions or removals.
- Counts: keep the selected total and matching total, while labeling unknown source-pool counts as unknown.
- Duplicates: record how overlapping clips and repeated views of one incident are handled.
These are suggested review fields, not a claim about Serve’s interface or export format. Ask a demonstration to show which selection information can be retained. Use a deliberately concentrated collection to investigate its intended cases, and use a separately defined measurement plan when the question is fleet frequency. No dataset was accessed or curation tool tested for this article.
Which collection does the percentage describe?
Primary sources
Serve Robotics: November 5 Physical AI Services announcement