Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

Built on Qwen3.5-4B, the model adds an external BEV perception head and a flow-matching trajectory planning module without altering the underlying VLM architecture.

• Performs 3D object detection, semantic occupancy prediction, BEV map segmentation, and visual question answering simultaneously
• Generates 5-second future ego trajectories at 10 Hz via a diffusion transformer Planning Expert, with optional text-reasoning conditioning
• Trained on 2.83 million public samples across unified cross-dataset trajectory formats to mitigate catastrophic forgetting
• Achieved an average driving QA score of 69.43 while retaining general multimodal reasoning performance across standard benchmarks

Autonomous driving researchers can adapt a single unmodified general-purpose VLM to handle end-to-end perception, reasoning, and trajectory planning.

Sources

• Qwen: Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving

Read this as text

Back to the AI news