Xiaomi-Robotics-1

Xiaomi-Robotics-1 addresses the scarcity of robot training data by combining 100,000 hours of embodiment-free pre-training data across diverse real-world scenarios with 7,200 hours of real-robot post-training data collected in homes. The two-stage approach—following language model training paradigms—demonstrates clean scaling behavior where validation performance improves as data and model size increase, suggesting robots can achieve greater capabilities through large-scale training similar to foundation models in language and vision.