Valid (Usable) Data

The portion of raw collected data that passes quality control and can actually be used to train a model.

This is a common term in the embodied-AI data-collection industry, without a single standardized definition. Some fraction of raw recorded data is always unusable: the task failed, the arm touched a prop it shouldn't have, the multiple sensor streams weren't time-synchronized, calibration was off, or there was occlusion or dropped frames. What remains after quality control filters these out is called valid (usable) data, and the fraction of raw data it represents is called the valid data rate (also called the pass rate). The head of the Beijing Humanoid Robot Innovation Center's data facility told media in April 2026 that the facility's pass rate had once been only about 50%, but stabilized above 95% after operator training and improvements to quality-control standards. The industry is also increasingly wary of judging data purely by hour count: frame rate and calibration have to meet a bar, scenes/objects/tasks need enough diversity, and ultimately the data has to pass client acceptance and prove itself in actual model performance.

ExampleA data collection worker records 100 demonstrations of “put the cup in the cabinet”; 8 have the cup fall, and 5 have the wrist camera drop frames. After quality control removes these, 87 remain as valid data — a valid data rate of 87%.

Also called

Valid Data Rate, Data Pass Rate

Related

Data Quality Control · Data Collector (Teleoperator) · Data Collection SOP · Data Cleaning · Data Curation · Embodied AI Training Ground (Robot Data Collection Center)

Sources

规模化「上课」为具身智能加注「数据燃料」（证券时报网，2026-04） (Chinese)
具身智能迈向 2.0：数据采集从训练场走向真实世界（科学网转澎湃新闻，2026-09） (Chinese)

As of

2026-09