Xiaomi Open-Sources Embodied AI Foundation Model Xiaomi-Robotics-1
Xiaomi has officially open-sourced its embodied AI foundation model, Xiaomi-Robotics-1, releasing its complete training pipeline, deployment code, and evaluation benchmarks. The model was pre-trained on over 100,000 hours of Universal Manipulation Interface (UMI) data and post-trained on over 10,000 hours of cross-embodiment data. This release provides the robotics and AI community with an "out-of-the-box" foundation model, accelerating research in cross-embodiment robot learning and physical task execution. By open-sourcing the entire pipeline, Xiaomi lowers the barrier to entry for developing adaptable, multi-purpose robotic systems. Xiaomi-Robotics-1 utilizes a two-stage training paradigm where the pre-training phase focuses on general action generation based on visual observations and language descriptions. The post-training phase utilizes cross-embodiment data to ensure the model can adapt to different physical robot configurations.
## BACKGROUND
Embodied AI refers to artificial intelligence embedded in physical entities, like robots, allowing them to interact with and learn from their physical environment. Universal Manipulation Interface (UMI) is a framework designed to transfer human demonstrations directly to deployable robot policies. Cross-embodiment learning is a technique that trains a single control policy to operate across different robot bodies with varying kinematics and sensors.