Xiaomi has officially open-sourced its robotics base model, named Xiaomi-Robotics-1. The announcement was made via the company’s official technology Weibo account on August 5. The release covers the entire process from real-device post-training to model deployment and includes benchmarking evaluation code.
Training Data and Model Design
Xiaomi-Robotics-1 was initially pre-trained on over 100,000 hours of UMI (Unstructured Manipulation Interaction) real-world data. Following this, it underwent more than 10,000 hours of post-training across different robot platforms to improve cross-embodiment capabilities. This two-stage training approach first equips the model with generic action generation skills and later adapts it to specific hardware embodiments.
Functionality and Availability
In the pre-training phase, given current visual input and language instructions, the model predicts a sequence of actions to shift a scene from its current to target state. This design aims to provide a ready-to-use foundation model for robotic manipulation tasks.
The open-source project including code and model checkpoints is available on Xiaomi Robotics’ official website, GitHub, and Hugging Face:
Context and Significance
This open-source effort follows Xiaomi’s July announcement of Xiaomi-Robotics-1 as a “plug-and-play” robotics base model. By combining extensive real-world data and cross-embodiment training, the model targets broad applicability in robotic control and manipulation tasks. The full release of training and deployment materials can support researchers and developers in robotics AI development.
Users interested in exploring the model and related tools can download the code and data using the links above. The availability of benchmarking scripts also facilitates performance comparison and evaluation on custom tasks or environments.
Source: ithome.com
