Xiaomi Open-Sources Embodied AI Foundation Model Xiaomi-Robotics-1 (4 min read)

Xiaomi has open-sourced its Embodied AI foundation model Xiaomi-Robotics-1. This release includes the entire pipeline from actual robot post-training to model deployment, along with related benchmark evaluation code.

Xiaomi-Robotics-1 was pre-trained on over 100,000 hours of UMI (Universal Manipulation Interface) data and underwent post-training on over 10,000 hours of cross-embodiment data. It aims to build general-purpose robot intelligence applicable to different robot hardware.

This release is significant in that it provides an end-to-end development blueprint at a time when most leading robot models are kept internal to companies. Developers and researchers can experiment with the model and evaluation tools via the project website, GitHub, and Hugging Face.

However, the model architecture, parameter count, and specific benchmark performance have not been disclosed. If the UMI data focuses on tabletop tasks and manipulation scenarios, domain generalization in real-world environments may be limited, requiring separate validation for initial adoption.

Xiaomi's open model contrasts with closed approaches by companies like Figure AI and Tesla, and aligns with open-source robotics projects such as LeRobot. Considering Xiaomi's humanoid robot CyberOne and China's manufacturing ecosystem, this release can also be interpreted as a strategic move to expand the research ecosystem and secure talent.