Xiaomi CEO Lei Jun has officially announced the release of Xiaomi-Robotics-0, the company’s first large-scale model built to help robots better understand and interact with the physical world. Unlike many firms that keep their robotics software proprietary, Xiaomi is making this system open source. Both the source code and trained models are now publicly available, allowing researchers and developers to deploy it across various robotic platforms.
Xiaomi-Robotics-0 adopts a dual-architecture design that separates high-level reasoning from motion control, building on the Qwen3 language model. The system was trained on a massive dataset that includes roughly 200 million distinct robot movements and more than 80 million image-text pairs. This extensive training enables robots to interpret complex instructions, recognize real-world objects, and execute precise physical actions.
>>>Battery replaces BW54 for XIAOMI Watch S2 46mm
One of the key breakthroughs addresses a common robotics limitation: the delay between planning and movement. With a technique known as a Lambda-shaped mask, the model can plan upcoming actions while completing current ones. This allows for smoother, more continuous motion while remaining responsive to new visual input. In simulation testing, Xiaomi-Robotics-0 achieved a success rate of nearly 99% on the LIBERO benchmark, outperforming many competing systems.
Beyond simulations, Xiaomi has demonstrated the model’s real-world capabilities. Robots powered by Xiaomi-Robotics-0 can dismantle complex Lego structures made of up to 20 bricks and fold towels with near-human dexterity. The system can even adjust dynamically—for example, flinging a towel to reveal a hidden corner or returning an extra towel if two are picked up by mistake.