Embodied AI data volume outpaces usable training gains, Leiphone reports

Leiphone (雷峰网) argues that embodied AI companies are collecting more robot training data, but usable data and gains in trained models have not kept pace.

The article divides the data pipeline into five stages: filtering, annotation, action conversion, verification, and feedback iteration. It names persistent problems in quality standards and in transferring learning across different robot bodies (cross-embodiment transfer). The report does not include measured figures for data volume or model performance, so the size of the gap is not quantified.