Xiaomi's new robotics release is not a chat model or a stage demo. Xiaomi-Robotics-1 asks a narrower industrial question: can a robot policy model improve predictably as data and model size grow?
The official project page describes it as "a ready-to-use robot foundation model trained on over 100K hours of real-world manipulation trajectories." In Chinese, the useful reading is simple: Xiaomi is trying to move robot learning from one-off task tuning toward a data factory.
The data bottleneck is the story
Robot data is expensive because it comes from bodies, operators, sites and maintenance. Xiaomi tries to loosen that constraint with embodiment-free UMI trajectories across more than 1,700 household, commercial, industrial and outdoor scenarios. A vision-language model splits long trajectories into clips and labels state changes, so the policy learns how actions change the scene rather than memorizing one robot's joint commands.
Post-training then connects that broad representation to real robots. Xiaomi reports more than 7,200 hours of in-house real-robot data from real homes, and Chinese reporting puts the full cross-embodiment post-training set at about 11,000 hours.
Strong benchmarks, visible caveats
The headline numbers are strong: 74.5% on RoboCasa, 57.4% on RoboCasa365, 59.1% on VLABench and 13.93% on RoboDojo. RoboCasa's independent entry lists 80.2% Atomic-Seen, 57.1% Composite-Seen and 32.1% Composite-Unseen, evaluated with private access to model weights and code.
For new-task adaptation, Xiaomi says less than 10 hours of demonstrations per task produced a 75% overall success rate, compared with 40% for the pi0.5 baseline; less than 40 hours lifted the result to 85% versus 53%.
That still is not a mass-production verdict. Public benchmarks do not measure long shifts, hardware wear, recovery after mistakes or household safety. GitHub also says code and model weights are still coming, while the RoboCasa entry marks the model as not open source. The next checks are code release, third-party replication and continuous real-world reliability.
Sources: Zhidongxi, Xiaomi Robotics project page, CocoLoop, GitHub repository, RoboCasa Leaderboard and CLS; checked 100K hours, 1,700+ scenarios, 7,200 real-robot hours, 75%/85% low-data adaptation, RoboCasa365 57.4% and RoboDojo 13.93% success rates.