Xiaomi has open-sourced Xiaomi-Robotics-1 (XR-1), a robot foundation model the company says outperforms rival systems on multiple manipulation benchmarks, according to the model’s GitHub repository published by the Xiaomi Robotics team.
The code and model checkpoints went live on GitHub on August 3, roughly three weeks after Xiaomi first described the system in a technical report posted to arXiv on July 16. The repository, released under the permissive Apache 2.0 license, includes tools for post-training, inference, deployment, and evaluation, with model weights also mirrored on Hugging Face and ModelScope.
What the model does
XR-1 is a vision-language-action (VLA) model, a class of AI that reads a camera feed and a written or spoken instruction, then outputs the physical actions a robot arm or hand should take. Xiaomi built it by combining a Qwen3-VL vision-language backbone with a diffusion-transformer action decoder in a mixture-of-transformers architecture.
The company trained the model in two stages. First, it pre-trained XR-1 on more than 100,000 hours of “embodiment-free” trajectories collected with a handheld data-capture rig called UMI (Universal Manipulation Interface), spanning over 1,700 real-world scenes in homes, shops, factories, and outdoor settings. It then fine-tuned the model on more than 10,000 hours of data gathered directly from robots, aligning the pre-trained skills with specific hardware and natural-language commands.
Benchmark results
According to the GitHub repository, XR-1 leads its next-best competitor by 2.6 percentage points on the RoboCasa benchmark (74.5%), by 23.2 points on RoboCasa365 (57.4% versus 46.6%), by 11.1 points on VLABench (59.1%), and by 58.3% relative on RoboDojo (13.93 versus 8.80). Xiaomi also says the model reaches a 75% task success rate after fewer than 10 hours of task-specific demonstrations, a sign it can be adapted to new jobs without extensive retraining.
Why it matters
Robotics researchers have increasingly argued that manipulation AI is held back less by model architecture than by the scarcity of real-world interaction data. By publishing both the training pipeline and the weights, Xiaomi is betting that a large, embodiment-agnostic dataset can generalize across different robots and tasks — and is inviting outside labs to test that claim directly rather than take its benchmark numbers on faith.
The release also extends a broader wave of Chinese firms open-sourcing frontier-adjacent AI models this year, from language models to now embodied systems, a trend that has been reshaping cost and access dynamics across the industry.