Key Takeaways
1. Xiaomi launches Xiaomi-Robotics-0, its first large-scale model for robots to interact with the physical environment.
2. The system is open-source, providing computer code and mathematical models for researchers and developers.
3. The model features a unique architecture that separates cognitive processing from movement control, trained on extensive datasets.
4. It allows robots to strategize actions while executing current tasks, achieving a nearly 99% success rate in simulations.
5. Real-world applications include complex tasks like disassembling Lego structures and folding towels with high precision.
Xiaomi’s CEO Lei Jun has made a big announcement about the launch of Xiaomi-Robotics-0, which is the company’s first large-scale model aimed at enabling robots to comprehend and engage with the physical environment. In contrast to numerous other firms that keep their robotic software secret, Xiaomi is opting to unveil this new system as an open-source initiative, allowing anyone to utilize it for creating improved robots. The release includes both the computer code and mathematical models, which are currently accessible for researchers and developers to download and apply on various types of robot hardware.
Innovative Design
The model features a unique architecture that divides tasks between a cognitive part for processing thoughts and a distinct section for movement control, building on the Qwen3 language model. The team trained this model utilizing a vast dataset encompassing around 200 million unique robot movements and over 80 million examples of images and text. Such thorough training enables the robot to interpret complex instructions and identify objects in the physical world while also meticulously planning its actions.
Enhanced Performance
This innovative model addresses a typical issue where robots hesitate to move as they think, as it allows the robot to strategize its next actions while still executing its current ones. Xiaomi employs a method known as a Lambda-shaped mask, which guarantees that the robot transitions smoothly based on prior actions while remaining prepared to respond instantly to new visual cues. Testing has revealed that this new system excels in computer simulations, achieving an impressive success rate of nearly 99 percent in the LIBERO benchmark, outshining many rival models.
Real-World Applications
Additionally, Xiaomi showcased the robot’s ability to carry out challenging tasks in real life, such as disassembling intricate Lego structures composed of as many as 20 bricks or folding towels with remarkable precision. The robot demonstrates intelligence by tossing a towel to uncover a hidden corner or even returning an extra towel if it mistakenly collects two at once.
Source:
Link