Xiaomi Open-Sources 38B World Model and Releases Robotics-1 VLA Foundation Model
Key Questions
What models has Xiaomi recently open-sourced for robotics?
Xiaomi released FlashAR+, a 38B world model for robotics data generation that delivers an 82× speedup. It also open-sourced Robotics-1, a VLA foundation model trained on over 100K hours of real-world UMI data.
How does Xiaomi Robotics-1 perform on robotics benchmarks?
Robotics-1 achieves SOTA results on RoboCasa365 (57.6%) and RoboDojo (20.07). It also shows strong few-shot fine-tuning performance, reaching 75% compared to π0.5's 40%.
What benefits do Xiaomi's open-source releases provide to robotics research?
The tools directly address the data bottleneck in robotics development. External validation shows U0-augmented data from the world model boosts π0.5's OOD success rate from 36.9% to 63.2%.
Xiaomi released FlashAR+, a 38B world model for robotics data generation, achieving 82× speedup. External validation shows U0-augmented data boosts π0.5's OOD success rate from 36.9% to 63.2%. Separately, Xiaomi Robotics-1 is a VLA foundation model trained on 100K+ hours of real-world UMI data, achieving SOTA on RoboCasa365 (57.6%) and RoboDojo (20.07) with strong few-shot fine-tuning (75% vs π0.5's 40%). These open-source tools directly attack the data bottleneck and set a new bar for generalist robot policies.