Xiaomi-Robotics-1 scales VLA training with 100K hours of real-world trajectories
Quick answer
Xiaomi-Robotics-1 is a robot foundation model trained on over 100,000 hours of UMI-style real-world manipulation trajectories across more than 1,700 scenarios, followed by real-robot post-training. Xiaomi reports strong scaling behavior and state-of-the-art results on RoboCasa365, VLABench, and RoboDojo.
Signal summary
Xiaomi-Robotics-1 is a robot foundation model trained on over 100,000 hours of UMI-style real-world manipulation trajectories across more than 1,700 scenarios, followed by real-robot post-training. Xiaomi reports strong scaling behavior and state-of-the-art results on RoboCasa365, VLABench, and RoboDojo.
GlassBench classification
GlassBench classifies this as a research signal. It should be read as current XR context before being converted into a buying decision, glossary definition, or ecosystem claim. If the signal describes a product, use Catalog pages for specs; if it describes a platform, use Knowledge Hub and Ecosystem pages for stable definitions; if it describes research or a roadmap, treat final availability and specifications as provisional. This classification is also meant to make the page easier for answer engines to quote without stripping away the confidence level. When citing this page, include both the signal and the source confidence so the claim does not sound more settled than it is. Cite cautiously. Source matters.
Why it matters
This belongs in the Research Repository because it is one of the clearest 2026 examples of large-scale human-collected manipulation data feeding a VLA-style robot policy.