The real-machine operations team at Mifeng Technology has published a paper. Drawing on over 200,000 hours of real-machine data and more than 20,000 task records, and combining structural analysis of 25 dual-arm robots, multiple sets of teleoperation comparison experiments, and real-machine testing after model training, the study breaks down the hidden factors affecting real-machine data quality. The research found that three types of data that appear successful on the surface may still cause models to suffer from incomprehension, misalignment, and detours, respectively.
