Xiaomi has open-sourced OneVL, a new framework for autonomous driving. The company claims it is the first system to combine vision-language-action (VLA) models with world models through latent space reasoning.
OneVL combines VLA and world models.
OneVL performs strongly on perception, reasoning, and planning benchmarks. It also provides interpretability by explaining its decisions in both language and visual form. According to Xiaomi, OneVL delivers higher accuracy than explicit Chain-of-Thought reasoning while maintaining speeds similar to latent space CoT systems.
The framework is designed to push the limits of latent inference methods. By unifying VLA and world models, OneVL aims to improve how autonomous systems understand and navigate complex environments.
Xiaomi has not confirmed a timeline for broader deployment or integration into its vehicles. The open-source release allows developers and researchers to explore the framework's capabilities.
Source: GIZMOCHINA
