Physical AIRoboticsAI02 sources

Xiaomi open-sources XR-1, a 100,000-hour manipulation model

A five-fingered anthropomorphic robotic gripping hand with a pale casing, held palm-forward and mounted on a mechanical wrist.

NearEMPTiness, via Wikimedia Commons (CC BY-SA 4.0)CC BY-SA

Xiaomi released the code and the model checkpoints for Xiaomi-Robotics-1 (XR-1) on 3 August 2026 under the Apache 2.0 licence, on GitHub and Hugging Face. XR-1 is a vision-language-action model for mobile manipulation: it couples a pre-trained Qwen3-VL vision-language backbone to a diffusion-transformer action head through a mixture-of-transformers architecture.

The training scale is what distinguishes it. According to the repository, the model was pre-trained on more than 100,000 hours of embodiment-free trajectories captured with UMI handheld rigs, then post-trained on more than 10,000 hours of cross-embodiment data spanning over 1,700 scenarios across household, commercial, industrial and outdoor environments.

The accompanying technical report was posted to arXiv on 16 July 2026 and has not been peer-reviewed. It reports a 57.4% success rate on the RoboCasa365 benchmark against a previous best of 46.6%, and an average score of 20.07 on RoboDojo against a prior 13.07. Both figures are the authors’ own, measured in simulation.

Sources

  1. [1]Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World TrajectoriesarXiv (Xiaomi Robotics Team)·
  2. [2]XiaomiRobotics/Xiaomi-Robotics-1GitHub·