Qwen-VLA unifies vision-language-action modeling across robots and tasks

Alibaba's Qwen team introduced a foundation model for manipulation, navigation, and trajectory prediction. Embodiment-aware prompts and a DiT-based action decoder helped it reach 97.9% on LIBERO, 73.7% on Simpler-WidowX, and 87.2% on RoboTwin-Hard. Its out-of-distribution success rate in real-world ALOHA tests was 76.9%. The paper did not specify parameter count or license.

Visit original source