Xiaomi has introduced Xiaomi-Robotics-0, its first large-scale robotics foundation model aimed at enabling machines to better understand and interact with the physical world, Chief Executive Lei Jun announced.
The company said the model, which contains 4.7 billion parameters, is being released as an open-source project, making its code and mathematical frameworks available to researchers and developers. Xiaomi described the move as an effort to support broader innovation in robotics by allowing the system to be deployed across different hardware platforms.
See also: Unitree Humanoid Robot Completes Extreme Cold Endurance Test Ahead of Winter Games
According to the company, Xiaomi-Robotics-0 is designed to standardize physical intelligence by separating high-level reasoning from low-level motion control. “By separating the robot’s thinking process from its movement control, the system eliminates common delays and allows for smooth, human-like actions in real-time,” the original announcement said.
The model builds on the Qwen3 language model and was trained using roughly 200 million distinct robot motion sequences alongside more than 80 million image and text samples. Xiaomi said this training enables robots to interpret complex instructions, recognize objects and plan physical actions with greater precision.
See also: China’s Robotics Sector Pushes Boundaries With Fast Humanoid Design
The company said internal testing showed the system achieved a success rate of nearly 99% on the LIBERO benchmark in simulation environments. Xiaomi also demonstrated real-world tasks including disassembling Lego structures of up to 20 bricks and folding towels, with the robot adjusting its actions if it mistakenly picked up more than one item.
Source: Xiaomi Robotics