Xiaomi-Robotics-1: больше данных важнее размера модели при обучении роботов

Xiaomi trained its Xiaomi-Robotics-1 model on more than 100,000 hours of motion data collected by humans using handheld grippers with cameras, rather than by robots themselves. In July 2026, the company showed that quality gains come primarily from data volume, not model size.

How the data was collected

The data for Xiaomi-Robotics-1 was recorded not by robots, but by humans: they picked up grippers with built-in cameras and performed ordinary household manipulations. This approach is cheaper and faster than teleoperating physical robots, and it made it possible to collect more than 100,000 hours of motion recordings.

• Model: Xiaomi-Robotics-1 from Xiaomi
• Training data volume — over 100,000 hours of movements
• Data was collected by humans using handheld grippers with cameras, not robots
• Key finding: growth in data matters more than growth in model size
• The absolute share of successfully completed tasks still remains low

Why data matters more than model size

Adding training data improved Xiaomi-Robotics-1's results noticeably more than increasing the number of model parameters did. This is a direct counterargument to the logic common for large language models, where the bet is often placed on scaling the network itself. In robotics, however, where what's actually lacking is recordings of real physical actions, the bottleneck turns out to be the volume and diversity of data, not the model's computational capacity.

More data delivers a bigger gain than increasing model size — the key finding of

Xiaomi's research on Xiaomi-Robotics-1.

Where the approach hits its limits

The quality gains for Xiaomi-Robotics-1 have not yet plateaued: the curve keeps rising as more data is added, which means the potential is not yet exhausted. But at the same time, absolute performance remains low — the model successfully completes only a portion of tasks. In other words, the data-scaling method works, but a reliable household robot that confidently handles manipulations is still a long way off.

What this means

Collecting data with human hands using cheap gripper-cameras could become a way to bypass robotics' main shortage — the lack of recordings of real actions. While absolute quality remains low, since the gains have not hit a ceiling, betting on data rather than model size sets the direction for the next generation of manipulator robots.

Frequently asked questions

How did Xiaomi collect data for Xiaomi-Robotics-1?

The data was recorded by humans: they performed household actions using handheld grippers with built-in cameras, rather than operating robots. This made it possible to collect more than 100,000 hours of motion recordings faster and cheaper than through physical robots.

What matters more for training a robot — data or model size?

According to Xiaomi's results, data. Adding training recordings improved Xiaomi-Robotics-1 significantly more than increasing model size did, and this gain has not yet plateaued.