Xiaomi Releases Embodied Foundation Model, First Systematic Verification of Robotics Scaling Law
Miles Bennett
Xiaomi released Xiaomi-Robotics-1, an embodied foundation model pre-trained on 100,000 hours of real-world manipulation data — the first systematic validation in China that scaling law holds for robot policy models. This means robot capability gains are shifting from manual tuning to a predictable 'more data, more ability' path.
Scaling law verified — what does that actually prove?
Scaling law — the principle that bigger models trained on more data get predictably better — was already well-established in large language models. In robot policy models, systematic evidence had been missing.
Xiaomi's experiments show: expanding pre-training data from 2,500 to 20,000 hours drives continuous drops in action-prediction loss; scaling parameters from 2 billion to 5 billion to 10 billion delivers steady improvement.
This means → robot capability now has a quantifiable growth path: invest more data and compute, get stronger performance — no longer reliant on engineers iterating by trial and error.
Where did 100,000 hours of data come from?
The data is not all from robots. Xiaomi built a proprietary portable device called UMI — Universal Manipulation Interface — that clips onto a human hand and records everyday actions in homes, offices, and factories.
To label this at scale, Xiaomi built an automated annotation pipeline using vision-language models: long trajectories are sliced into segments, each tagged with a state-change description.
In plain terms = the model doesn't learn "copy the human's arm movement." It learns "understand how an object goes from state A to state B" — the underlying physics, not the surface motion.
What is the "pre-train + post-train" two-stage design?
Stage one (pre-training): 100,000 hours of general data teach the model foundational physical skills — grasping, sorting, manipulating.
Stage two (post-training): roughly 11,000 hours of cross-embodiment data — covering mobile manipulators, dual-arm robots, plus public datasets like Bridge V2, RT-1, and DROID — align the model to different robot bodies and natural-language instructions.
This means → the model works "out of the box": in unseen real home environments, it can follow voice commands to organize a shoe cabinet or tidy a desk, with no per-scene retraining.
How strong are the benchmark results?
On the RoboDojo simulation benchmark, Xiaomi-Robotics-1 scored 20.07 with a 13.93% success rate — a clear lead over the previous best of 13.07 and 8.80%.
On RoboCasa365, which covers hundreds of real household scenarios, it posted a 57.4% average success rate, surpassing the 46.6% record held by Google-affiliated teams.
For complex tasks, the model needs on average fewer than 10 hours of fine-tuning data to far outperform a model trained from scratch. This reflects genuine transfer: the general capabilities built during pre-training carry over to new scenarios.
Where does this sit in Xiaomi's robotics strategy?
This is the finale of a three-day robotics blitz from July 14 to 16: day one covered robot hardware entering factory internships, day two unveiled a unified generative model, and day three delivered this foundation model.
Together, the three releases sketch a "hardware — data — model" technology closed loop.
In plain terms = Xiaomi's narrative is: the robot body is ready, the data pipeline is open, and the brain is now in place. The next test is whether this system can deliver on scaling law's promise at real-world, large-scale deployment.
Content is for reference only, not financial advice.