Xiaomi open-sources a robot model trained on 100,000 real hours

Xiaomi open-sourced Xiaomi-Robotics-1, a robot model that beats five rival systems on a standard manipulation benchmark, per the arXiv paper.

The five it beat count among the strongest systems in robot learning. An open, free model that outperforms all of them resets the baseline every other lab has to clear.

The model scored a 74.5% average success rate on RoboCasa, ahead of RLDX-1, Cosmos Policy, GR00T N1.6, Pi-0.5, and Pi-0-FAST. On the harder RoboCasa365 suite, it hit 57.6%, a new state of the art over the prior best of 46.6%.

Xiaomi-Robotics-1 doesn't use a new architecture. Its edge comes from the training set: 100,000 hours of real manipulation trajectories, collected from physical robots instead of simulation.

That kind of real-world data is expensive and slow to collect at scale. That's exactly why it works as a barrier other labs can't easily clear.

Robot learning's next round of benchmarks will likely go to whoever has the most real-world manipulation hours, not whoever designs the cleverest model. Worth watching whether Xiaomi keeps releasing weights at this scale, or whether this is the high point before the real advantage gets held back.

The arXiv paper's Hacker News post drew 342 points in nine hours, fast for a robotics paper on a site that usually favors software.