Qwen-Drive-1.0 folds 3D perception and planning into one VLM

One model, three jobs: 3D perception, visual question answering, and motion planning, all inside a single vision-language foundation. That is the pitch of Qwen-Drive-1.0, which its authors describe as "an initial step towards a vision-language foundation model for autonomous driving."

The system keeps the pretrained VLM architecture intact and adds two driving-specific components. An external bird's-eye-view perception head jointly performs "3D object detection, semantic occupancy prediction, and BEV map segmentation," and a Planning Expert conditions on the shared VLM representations to "generate future ego trajectories." A staged training recipe combines driving supervision with general-purpose vision-language data, meant to acquire driving competence "while helping preserve broad visual understanding and instruction-following capabilities."

The abstract publishes no benchmark numbers. It reports "strong 3D perception and driving scene understanding" and "highly competitive motion-planning performance" across open-loop, pseudo-closed-loop, and closed-loop settings, without naming the baselines or the margins.