Xiaomi Opens Source Xiaomi OneVL Autonomous Driving Model, First in Industry to Unify Multiple Tech Routes

In the autonomous driving field, VLA and world models used to be separate technical routes: VLA focuses on scene understanding and driving action output, while world models emphasize future scene prediction. Xiaomi's Xiaomi OneVL framework unifies them via latent space reasoning for the first time, achieving SOTA in multiple benchmarks like ROADWork.

The model not only improves reasoning speed and accuracy but also explains decisions from both language and visual dimensions—using text to justify driving choices and predictive images to show subsequent scenes. Xiaomi has fully open-sourced its model weights and training/inference code for developers to use.