Mifeng Releases 200,000-Hour Real-World Robot Data Quality Study
Leaderobot reports that embodied AI is drawing on an increasingly diverse range of data sources, but the key question is whether models actually learn useful lessons from all of it. Mifeng has released a study of data quality covering 200,000 hours of real-world robot data, focused on the “effective experience” used in post-training embodied agents. Rather than treating data volume as the goal, the work examines how real-machine data can be screened and organized for training value. The study is positioned as a reference for teams building embodied agents, highlighting that post-training gains depend on data quality as much as scale. The available abstract does not disclose detailed benchmark results, but frames data quality as a central bottleneck for embodied AI.