Skip to main content

Xiaomi unveils robot brain trained on 100,000 hours of real-world data

Xiaomi has taken a big leap into the world of physical AI. On July 16, the company officially unveiled Xiaomi-Robotics-1, an embodied foundation model designed to handle real-world mobile tasks. Think of it as a brain for robots that can actually get up and do things.

What makes this model special? It was pre-trained on a massive dataset: 100,000 hours of real-world trajectories. That's not simulated data—it's actual movements collected from robots doing tasks in homes, offices, and factories. The team used a device called UMI (Universal Manipulation Interface) to gather this data, then combined it with a visual language model to automatically annotate everything in just two weeks.

Image

But the training didn't stop there. In a second phase, the team used about 10,000 hours of cross-body data—meaning data from different robot types—to align the model's understanding of instructions and actions. The result? A model that can work across multiple robot platforms without needing extensive customization. Xiaomi calls it "out-of-the-box" capability.

Xiaomi-Robotics-1 comes in three sizes: 2 billion, 5 billion, and 10 billion parameters. And here's the exciting part: as the model size and training data increased, performance kept improving. That's the "scaling law" in action—a principle that has driven breakthroughs in language models but is still being explored in robotics. The model set new state-of-the-art records on benchmarks like RoboCasa365 and RoboDojo.

Image

This isn't just another research paper. Xiaomi is showing a clear path: start with massive pre-training on real-world data, then fine-tune with cross-body data, and finally adapt with a small amount of task-specific data. It's a recipe that could help robots move from lab demos to real-world applications.

For anyone following AI, this feels like a milestone. Physical AI has long struggled with the "data problem"—robots need huge amounts of diverse data to learn, but collecting that data is slow and expensive. Xiaomi's approach suggests that with the right tools and scale, those barriers can be broken.

Key Points

  • Xiaomi launched Xiaomi-Robotics-1, an embodied foundation model for mobile robots
  • Pre-trained on 100,000 hours of real-world data from home, commercial, and industrial settings
  • Post-trained with 10,000 hours of cross-body data for multi-platform compatibility
  • Available in 2B, 5B, and 10B parameter versions, showing clear scaling benefits
  • Set new state-of-the-art results on RoboCasa365 and RoboDojo benchmarks