The model uses a phased training approach that combines driving supervision with general visual language data. This allows the system to acquire driving-specific skills while maintaining general visual understanding and instruction-following capabilities.
Qwen-Drive-1.0-4B includes both SFT and RL planning experts. The official evaluation covers 3D perception, driving scene understanding, general visual language abilities, and various motion planning modes, including open-loop, pseudo-closed-loop, and closed-loop.
