Qwen-Drive 1.0 tells you why it brakes, just don't expect the explanation to match the maneuver
decoder ·
Alibaba's Qwen-Drive 1.0 combines spatial perception, traffic question answering, and route planning in a single AI model. Built on the Qwen3.5-4B language model, it adds two components: one that constructs a bird's-eye-view map of the surroundings and another that plans the vehicle's route. The researchers found that a text-image model does not automatically understand 3D space, so they deliberately trained the vision-language model on spatial tasks, which significantly improved performance. In simulator tests, retraining reduced the rate at which the car veered off the road from 24 percent to 12 percent, though it also drove more cautiously. The model's explanations for its driving decisions do not always match its actual maneuvers, and it can conflate situations such as a distant red light with a child stepping into the road. Performance also drops when processing footage from vehicles with different camera setups due to a lack of suitable training data. Qwen-Drive 1.0 is being released for free to the research community on Hugging Face, ModelScope, and GitHub.
decoder