Xiaomi-Robotics-1 Advances Robot Learning with Large-Scale Pre-Training

Jul 20, 2026

AI Summary

Xiaomi has developed Xiaomi-Robotics-1, a robot model that combines extensive embodiment-free pre-training with real-robot data to improve its capabilities. The model shows promising results in learning new tasks efficiently and achieving high success rates in both real-world and simulation environments.

Xiaomi-Robotics-1 utilizes 100,000 hours of embodiment-free pre-training data across over 1,700 scenarios, including household and industrial tasks.
An automatic annotation pipeline was created to label data segments with language descriptions, facilitating large-scale training.
The model undergoes two training stages: pre-training for general action generation and post-training for aligning actions with real robot capabilities and natural language instructions.
Post-training results indicate that as pre-training data and model size increase, the success rate of real-robot tasks also improves predictably.
Xiaomi-Robotics-1 can learn new tasks efficiently, achieving a 75% success rate with under 10 hours of demonstrations per task, and 85% with under 40 hours.
The model has achieved state-of-the-art results on four mainstream simulation benchmarks, demonstrating its generalization capabilities.
Overall, Xiaomi-Robotics-1 represents a significant advancement in scaling robot foundation models, effectively addressing the data scarcity challenge in robotics.

xiaomiroboticstechnologyinnovationautomation

Xiaomi-Robotics-1 Advances Robot Learning with Large-Scale Pre-Training

Related Stories

Workers Use Motion Trackers and Cameras to Train AI on Human Movement

Robotics Firms Use Human Workers' Videos to Enhance AI Training

Unitree Robotics IPO Sees 5,526 Times Subscription from Retail Investors

FCC's Ban on Chinese Robots Highlights Need for U.S. Manufacturing Strategy