Xiaomi's Robot Brain Ate 100,000 Hours of Real Hands
Xiaomi Robotics open-sourced Xiaomi-Robotics-1, a vision-language-action model that's basically a foundation model for a robot body. You give it a language instruction, it performs mobile manipulation tasks in environments it has never seen, out of the box, and it fine-tunes to new tasks with barely any data. The thing that makes it work is the scale of what it ate: over 100,000 hours of real-world manipulation trajectories.
Those trajectories were collected with UMI devices, the handheld grippers that let you record a demonstration without a robot in the loop, across more than 1,700 scenarios: households, shops, factories, outdoor spaces. That embodiment-free trick is how you get to 100,000 hours without buying 100,000 robots. Then they built an auto-labeling pipeline that annotates each clip with natural language describing how the scene changes, giving the model rich conditioning for what its actions are supposed to accomplish.
Why care if you're not building robots? Because this is the same playbook that gave us language models, pointed at the physical world. Scrape a mountain of demonstrations, learn a general policy, adapt cheaply downstream. The agents we cover mostly live inside a terminal; VLA models like this are the bridge to agents that have hands. Xiaomi shipping it open, weights and code, is the part that moves the field. Repo at github.com/XiaomiRobotics/Xiaomi-Robotics-1, paper at arXiv 2607.15330.
← Back to all articles
Those trajectories were collected with UMI devices, the handheld grippers that let you record a demonstration without a robot in the loop, across more than 1,700 scenarios: households, shops, factories, outdoor spaces. That embodiment-free trick is how you get to 100,000 hours without buying 100,000 robots. Then they built an auto-labeling pipeline that annotates each clip with natural language describing how the scene changes, giving the model rich conditioning for what its actions are supposed to accomplish.
Why care if you're not building robots? Because this is the same playbook that gave us language models, pointed at the physical world. Scrape a mountain of demonstrations, learn a general policy, adapt cheaply downstream. The agents we cover mostly live inside a terminal; VLA models like this are the bridge to agents that have hands. Xiaomi shipping it open, weights and code, is the part that moves the field. Repo at github.com/XiaomiRobotics/Xiaomi-Robotics-1, paper at arXiv 2607.15330.
Comments