A 38-billion-parameter open-source embodied AI model unifies scene generation, robot transfer, video generation, and image editing while improving robotics training efficiency.
Xiaomi has introduced Xiaomi Robotics-U0, a 38-billion-parameter multimodal embodied AI foundation model that combines four robotics tasks within a single architecture. The development is fully open-sourced with the release of model weights, source code, and other related resources allowing developers to build customised embodied AI applications.
The model combines four core capabilities. It generates robot training scenes from the text prompts, transfers robot trajectories to new environments while preserving their motion, generates robot interaction video sequences from observation and task instructions, and supports text-to-image generation and image editing task for robotics datasets. Overall, all of these capabilities combined offer a unified process to generate and enhance robot training data.
As per the information provided by Xiaomi, the model achieved the highest score among 126 participating models on the WorldArena benchmark test. With real robot testing performed on unknown lighting and background conditions, training with the generated data improved strategy task completion rates by an average of 26 per cent.
In order to optimise deployment speed, the model uses UNIS, an inference acceleration framework developed by Xiaomi that is claimed to be around 83 times faster than conventional autoregressive approach. It allows embodied AI training data generation on a large scale.
The open-source release includes the project website, GitHub repository, Hugging Face model weights, and ModelScope collection. According to Xiaomi, the release supports research and development in the field of embodied AI through a complete framework for robot data generation, training, and deployment.
