Xiaomi CEO Lei Jun has released Xiaomi-Robotics-0, the company's first large language model designed to help robots understand and interact with the physical world. Unlike most robotics companies that keep their AI systems proprietary, Xiaomi is releasing this 4.7 billion parameter model as an open-source project, making the code and mathematical models available for researchers and developers to use on different robot hardware platforms.
The model uses a split architecture that separates cognitive reasoning from movement control, building on the Qwen3 language model foundation. Xiaomi trained the system using approximately 200 million distinct robot movements and over 80 million general images and text examples, enabling it to understand complex instructions, recognize real-world objects, and plan physical actions with precision. The team employed a Lambda-shaped mask technique that allows robots to plan their next moves while still executing current ones, eliminating the common pause-and-think delay that plagues many robotic systems.
In benchmark testing, Xiaomi-Robotics-0 achieved a 99 percent success rate in the LIBERO simulation benchmark, significantly outperforming competing models. Real-world demonstrations showed the system handling intricate tasks like disassembling Lego structures with up to 20 bricks and folding towels with human-like dexterity. The robot displayed adaptive behavior during towel folding, such as flinging fabric to locate hidden corners or returning an extra towel if it accidentally grabbed two at once.
The open-source release includes full access to the model weights and training code, positioning Xiaomi as a major contributor to collaborative robotics development. The project is built on the HuggingFace Transformers ecosystem with Apache License 2.0, allowing deployment on standard Linux systems using Python 3.12, PyTorch 2.8.0, and consumer-grade GPUs with Flash Attention 2 and bfloat16 precision. Installation requires basic apt-get packages (libegl1, libgl1, libgles2) and works with Conda environments, making it accessible to developers running Ubuntu or other mainstream distributions. By making this technology freely available, the company is enabling broader experimentation with physical AI systems across academic and commercial applications.
