Xiaomi Open-Sources Robot Foundation Model XR-1, Trained on 100,000 Hours of Real-World Data

Bot Telegraph, July 22nd – The Xiaomi robotics team recently released the embodied intelligence foundation model Xiaomi-Robotics-1 (abbreviated as XR-1). This model was pre-trained on over 100,000 hours of real-world operational trajectories, covering more than 1,700 scenarios including homes, commercial spaces, industrial sites, offices, and outdoor environments.

XR-1 employs a two-stage “pre-training + post-training” approach: first, it learns general action generation capabilities using large-scale operational data collected by UMI devices. Subsequently, it uses approximately 10,000 hours of cross-embodiment data to adapt the model to real robots, enabling it to perform mobile manipulation tasks such as sofa tidying, tableware storage, and shoe cabinet organization based on natural language instructions.

Results published by Xiaomi show that XR-1 achieved leading performance across four simulation benchmarks: RoboCasa, RoboCasa365, VLABench, and RoboDojo. Specifically, it attained an average success rate of 57.4% on RoboCasa365 and 13.93% on RoboDojo. For new task adaptation, after fine-tuning the model with an average of less than 10 hours of demonstration data, the overall success rate across four real-robot tasks reached 75%.

As the world’s first unified generative model in the embodied field, Xiaomi-Robotics-U0 has simultaneously validated generation quality, data efficiency, and engineering efficiency.

Source: Bot Telegraph China