Xiaomi open-sources its multimodal autoregressive embodied generative base model, Xiaomi-Robotics-U0.

Xiaomi has open-sourced Xiaomi-Robotics-U0, a multimodal autoregressive embodied generation model with 38 billion parameters. This model is the first in the embodied generation field to uniformly cover four tasks: scene generation, trajectory transfer, interactive video, and general image editing. The model achieves precise and controllable generation through five-dimensional decoupled structured control and is equipped with a FlashAR+ inference acceleration solution, improving efficiency by nearly 83 times.