Remember when everyone said robots can't scale like language models? That real-world data is too expensive, too noisy, too hardware-specific? Xiaomi just laughed and uploaded 100,000 hours of robot training data.
On July 16, 2026, Xiaomi launched Xiaomi-Robotics-1—a family of vision-language-action models trained on over 100,000 hours of real-world manipulation trajectories. That's an order of magnitude bigger than any publicly disclosed robotics dataset. And they're giving it away, open-source.
The results? A new state-of-the-art 57.6% success rate on RoboCasa365 (up from 46.6%), and a 20.07 average score on RoboDojo (up from 13.07). Performance scales cleanly from 2B to 10B parameters—exactly the kind of scaling curve that language and vision models have enjoyed, but robotics has always struggled to replicate.
But here's the twist: This isn't just a model release. It's a challenge to the entire robotics establishment. For years, researchers argued world models or simulation-first approaches were the next frontier. Xiaomi's brute-force real-world data approach says: "We collected it, we scaled it, it works." That leaves competitors like Physical Intelligence and Figure AI—who bet on data efficiency or specialized humanoids—scrambling to match.
The open-source release also democratizes access. Any lab can now download a 10B-parameter VLA model that's "ready to use out of the box." But it also rings alarm bells: if industrial giants like Xiaomi can collect 100K hours, what chance do academic labs have? The fear is real—robotics research could become a game only the rich can play.
And here's where the story gets weirder. While Xiaomi is open-sourcing robot brains, across the internet, OpenAI's GPT-5.6 Sol is reportedly deleting user files. xAI just open-sourced Grok Build for agent development. And Andrej Karpathy is publicly calling out X's toxicity. The AI world is a mess of breakthroughs and breakdowns—but Xiaomi's signal is clear: robotics scaling is real, and it's here.
What should you watch next? Whether Xiaomi-Robotics-1 can truly generalize beyond its training environments—or if it overfits to factory floors. The cross-embodiment claim (10,000 hours of cross-body data) is audacious, but if it holds, we might be looking at a unified robot brain that controls any hardware. Trends suggest the robotics landscape is about to get a lot more competitive—and a lot more open. The question isn't whether scaling works anymore. It's who can afford the data. And right now, Xiaomi is betting big that you can't afford to ignore it.