Qwen-Drive-1.0 (huggingface.co)

The Qwen-Drive-1.0 has been launched, marking a significant advance in vision-language models specifically tailored for autonomous driving. This comprehensive framework combines 3D perception, visual question answering, and motion planning without altering the underlying Qwen3.5 architecture. It features an innovative BEV perception head that facilitates 3D object detection and map segmentation while a specially designed Planning Expert generates future trajectories through flow matching. This model utilizes a staged training approach that fuses driving and general vision-language datasets, enhancing driving capabilities while maintaining broad visual understanding. This development is crucial for the AI/ML community as it represents the first unified model that supports 3D perception and driving tasks at the pretraining stage, allowing for enhanced interaction with complex driving environments. Notably, Qwen-Drive-1.0 exhibits competitive performance across various evaluations, leveraging reward optimization techniques to improve decision-making and safety in closed-loop driving scenarios. Its architecture supports real-time perception and planning, setting the stage for more sophisticated applications in autonomous vehicles and reshaping the standards for future vision-language models in autonomous navigation.