Alibaba releases Qwen-Drive 1.0 driving AI model

Existing driving models typically fine-tune a general text-image model on traffic Q&A data. The researchers argue this has two weaknesses: such models still cannot reliably detect distances and positions, and they can suffer from "catastrophic forgetting," losing the broad general knowledge that is crucial in rare, unexpected situations. Qwen-Drive 1.0 is built to address both problems by adding a module that builds a bird's-eye-view map and a Planning Expert module, while running all driving functions through a shared language model.

The team found that training only the added component did not improve spatial accuracy; only when the vision-language model itself was trained on spatial tasks did performance improve significantly. This confirms that spatial understanding must be built in deliberately. The model is intended for modern vehicles where the infotainment system and driving system converge on a single computing unit, so a model that retains general knowledge is needed for cockpit tasks.

The stakes hinge on how well the model handles real-world driving beyond the researchers' own test procedures. The team notes that the model's explanations don't always match its planned maneuver and that it detects far less when processing footage from other vehicles with different camera setups. The release of the model to the research community could help address these gaps, though suitable training data for other camera configurations is still missing.