Xiaomi Robots Validate the Scaling Law

卡尔的AI沃茨 · wechat · 2026-07-16

This article systematically breaks down Xiaomi's newly released embodied robot tech stack, focusing on two models: Robotics-1 and Robotics-U0.

Key Conclusions

The author believes Xiaomi has validated the scaling law in robotics: capabilities consistently and steadily improve as data volume and model size increase.
In the past, robot data collection was expensive and hard to scale. Using portable capture devices like the UMI (Universal Manipulation Interface), Xiaomi scaled real-world operation data from around 10,000 hours to the 100,000-hour level.
The article notes that as data increased from 2.5K, 5K, 10K to 20K hours, action prediction loss consistently dropped and training stabilized. Scaling model size from 2B and 5B to 10B also showed a "bigger is better" trend.

Robotics-1

Acting as the second-generation brain, it focuses on manipulation capabilities in real-world environments.
Tested in unseen environments on tasks like shoe cabinet organizing, backpack packing, and desk tidying.
The article claims it ranks first on four mainstream benchmarks, including:
RoboDojo: 20.07 points
RoboCasa365: 57.4% average success rate
RoboCasa: 76.1%
VLABench: 59.1%
It also mentions superior few-shot adaptation compared to Physical Intelligence's Pi-0.5: requiring less than 10 hours of data per new task for fine-tuning, beating the latter across four new tasks.

Robotics-U0

This is a 38B parameter unified embodied synthesis model responsible for "imagination" and data augmentation.
It generates multi-view, geometrically consistent images of robot manipulation scenes from text, and transfers real trajectories into different environments.
The author suggests it can "expand" a few real trajectories into numerous varied scenarios to augment training data.
Training on data generated by this model boosted Pi-0.5's OOD success rate on real robot tasks from 36.9% to 63.2%.

Article Insights

Xiaomi's complete flywheel is: U0 generates data → Robotics-1 gets stronger → more real-world scenario data feeds back to U0.
The author views this as a crucial step toward out-of-the-box robotics, bringing us closer to household robot butlers.

More from Embodied