Xiaomi-Robotics-1 arrives as a ready-to-use robot foundation model trained on 100,000 hours
Trained on more than 100,000 hours of real-world manipulation trajectories and combining embodiment-free pre-training with real-robot data, Xiaomi-Robotics-1 is offered as a downstream-ready foundation model. Robotics has lacked the shared starting point that language modelling has had for years.
The bottleneck in robot learning has never been architecture, it has been data. Manipulation trajectories must be collected on physical hardware in real time, which cannot be scraped and cannot be parallelised the way text can. A hundred thousand hours represents an enormous capital commitment, and releasing a model trained on it transfers that cost to everyone downstream.
The hybrid recipe is the interesting design choice. Embodiment-free pre-training learns representations that do not assume a specific body; real-robot data grounds them in one. That combination is aimed squarely at the field's transfer problem, where policies trained on one platform historically fail on another.
If it works as described, it does for robotics what pre-trained language models did for NLP: turn every project from a data-collection campaign into a fine-tuning exercise. That is the change that moves a field from labs to products, and it is worth watching whether independent groups reproduce the transfer claims on hardware Xiaomi did not choose.
Radical Data Science — AI news briefs bulletin board for August 2026 → · NVIDIA — National Robotics Week — latest physical AI research →