Mifeng Study Finds Real-Robot Data Quality Gaps in Embodied AI Training
Mifeng Technology has released a study on real-robot data quality for embodied AI, based on more than 200,000 hours of real-robot data and over 20,000 task records. The project also analyzed the structures of 25 dual-arm robot platforms and ran post-training tests on physical robots to assess how data quality affects embodied model performance. The research aims to clarify which data properties matter most when scaling real-robot training.
The team reported that models retain about 90% of the extra paths contained in demonstrations. When detour data increased from 0% to 80% of the training mix, task completion time rose from 25.3 seconds to 32.6 seconds, indicating that inefficient demonstration trajectories can be learned along with successful task outcomes.
In a separate experiment, completing the visual information required for robot decision-making raised the accuracy of selecting the correct robotic arm from roughly 50% to 100%. This suggests that missing observations can severely limit policy performance even when the task itself is otherwise learnable. Visual completeness is therefore a key variable in data collection.
The study concludes that task success alone does not mean the underlying data is high quality. Observation completeness, consistency between data-collection and deployment calibration, and action efficiency all influence post-training results. Mifeng says the findings offer practical criteria for evaluating real-robot data before it is used to train embodied AI models.