Xiaomi-Robotics-1 Breaks Data Barriers with 100,000 Hours of Embodiment-Free Training
Xiaomi-Robotics-1
We introduce Xiaomi-Robotics-1, a foundation model that overcomes robotics data scarcity by pre-training on 100,000 hours of embodiment-free trajectories. By combining this massive scale with real-robot post-training, we demonstrate clear scaling laws where larger models achieve higher success rates in unseen environments. Our approach enables robots to learn complex tasks from minimal data while setting new state-of-the-art benchmarks.
"The scarcity of large-scale, high-quality data, more than anything else, has capped how far policy models could scale."
HN discussion
- Critics argue that robotics demos often rely on controlled environments and sped-up footage, noting that while an 80% success rate is achievable, covering the final 20% of reliability can take an entire lifespan.
- One commenter highlights the irony that open weights models distributed altruistically are proving more effective than proprietary systems, urging the US to release AI weights to prevent China from closing its technology sector once the competitive foil disappears.
- Skeptics point out that Xiaomi's vacuum technology already delivers reliable, competent results for floor cleaning, questioning the necessity of a general-purpose robot when specialized devices like the Roborock already outperform competitors like iRobot.
- Optimists compare current robotics breakthroughs to the Model T from Ford, suggesting that despite current limitations, the technology will evolve rapidly over the next decade or century.
- Concerns are raised that mass labor displacement by robots could lead to a society where a small cadre of elites own the machines while the rest of the population becomes serfs, unless strong social services and UBI are implemented.