Alibaba releases Qwen-Drive 1.0 AI model with spatial awareness limitations

Alibaba has released Qwen-Drive 1.0, an open-source AI model integrating spatial perception, traffic Q&A, and route planning. Built on Qwen3.5-4B, the model outperforms specialized systems in benchmarks and reduces simulator road-veering from 24% to 12% via reinforcement learning. However, researchers highlight spatial awareness limitations, noting that its driving maneuvers do not always match its textual explanations, and performance drops significantly on unfamiliar camera configurations.

Qwen-Drive 1.0 architecture

• ▪The perception module in Alibaba's Qwen-Drive 1.0 generates a bird's-eye-view map of the surroundings by identifying objects in three-dimensional space.
• ▪Alibaba's Qwen-Drive 1.0 AI model integrates three tasks into a single model: spatial environment perception, traffic question-answering, and route planning.
• ▪Alibaba released the Qwen-Drive 1.0 model for free to the research community on Hugging Face, ModelScope, and GitHub on September 7, 2026.
• ▪Alibaba's Qwen-Drive 1.0 is built on the Qwen3.5-4B base model and incorporates two additional components: a 3D perception module and a Planning Expert.

Spatial understanding limitations

• ▪The planned driving maneuvers of Alibaba's Qwen-Drive 1.0 do not always align with the textual reasoning the model provides beforehand.
• ▪Alibaba researchers confirmed that vision-language models capable of detailed image description do not automatically comprehend three-dimensional spatial relationships.
• ▪Alibaba's Qwen-Drive 1.0 exhibits spatial limitations, as its generated driving explanations do not always accurately identify the actual cause of a traffic situation.

Training methodology

• ▪The training process for Alibaba's Qwen-Drive 1.0 progresses sequentially from the perception module, to combined perception and question-answering, and finally to route planning.
• ▪Alibaba trained the vision-language component of Qwen-Drive 1.0 using a standardized compilation of 24 publicly available traffic scene datasets.

Benchmark performance results

• ▪In benchmark tests, Alibaba's Qwen-Drive 1.0 outperformed specialized driving models in most driving and perception categories.
• ▪In simulator testing, a version of Alibaba's Qwen-Drive 1.0 refined with reinforcement learning rewards reduced its road-veering rate from 24 percent to 12 percent.

Real-world testing gaps

• ▪Alibaba researchers noted that some Qwen-Drive 1.0 evaluation metrics rely on custom-built test procedures, limiting their ability to predict real-world driving performance.
• ▪Alibaba's Qwen-Drive 1.0 experiences a significant drop in spatial detection performance when processing footage from vehicles with unfamiliar camera configurations.

Security vulnerabilities

• ▪University of California, Santa Cruz researchers hijacked the DriveLM driving system, forcing it to swerve toward pedestrians by placing a labeled sign in the camera's view.
• ▪Researchers warn that pairing a language model with a driving function introduces new security attack surfaces to autonomous vehicle systems.