Qwen-Drive 1.0 combines perception, traffic answers and route planning
Alibaba’s research division has introduced Qwen-Drive 1.0, an AI model designed to perceive road scenes, answer traffic-related questions and plan a vehicle’s route within one system. It is built on Qwen3.5-4B and adds modules for constructing a 3D environment map and planning near-term movements. The same language model can therefore support both driving functions and cockpit interactions.
The researchers found that image description ability does not automatically provide three-dimensional understanding. Training the added components alone produced little spatial improvement; the underlying vision-language model also had to be trained on spatial tasks. Training combined 24 public traffic-scene datasets with standardized questions and answers, followed by route-planning training and reinforcement learning.
In simulations, reinforcement-trained versions reduced the rate of cars leaving the road from 24 percent to 12 percent, although they drove more cautiously and covered less distance. The system’s explanations were not always consistent with its actual decisions: it could confuse a distant red light with a child entering the road, for example. Performance also fell on footage from vehicles with different camera configurations. Qwen is making the model available to researchers through Hugging Face, ModelScope and GitHub.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.