Alibaba researchers developed Qwen-Drive 1.0, an AI model designed to handle spatial perception, traffic Q&A, and route planning. The system is based on the Qwen3.5-4B model, which was released in February. It incorporates two additional components: a bird's-eye-view mapper for 3D object detection and a Planning Expert for movement strategy.
The training process involves three stages: perception, combined perception and Q&A, and route planning, followed by reinforcement learning. The team utilized 24 public traffic datasets, which were standardized by an AI to ensure consistency. The model aims to prevent catastrophic forgetting, where specialized training causes a loss of general knowledge.
In simulator tests, reinforcement learning reduced the rate of vehicles veering off the road from 24 percent to 12 percent. While the model outperforms the base Qwen3.5-4B in traffic-related reasoning, researchers noted that its explanations do not always accurately reflect the actual causes of driving maneuvers.
The model is available for the research community on Hugging Face, ModelScope, and GitHub. The authors emphasize that integrating driving and infotainment systems requires models that maintain general capabilities alongside specialized driving performance.
