DATA BEATS SIZE IN ROBOT TRAINING, XIAOMI FINDS

Xiaomi's approach diverges from industry trends favoring larger language and vision models. Rather than scaling up architecture, the company prioritized data collection, using human operators with handheld grippers to generate training footage instead of relying on robotic data. Results show significant performance improvements from additional data inputs, with gains continuing as datasets expand. However, absolute success rates remain low, indicating the technology is early-stage. The findings suggest a different path forward for robotics AI development. While large foundation models dominate other sectors, robot control may benefit more from quantity and quality of motion data than parameter count. Xiaomi plans to continue scaling its dataset, suggesting confidence in the data-first methodology. The approach could reshape how companies train physical AI systems, shifting focus from model size to comprehensive behavioral datasets.