The system retains the pretrained Qwen3.5-4B architecture and attaches an external bird-eye-view perception head and a flow-matching Planning Expert. Staged training combines driving supervision with general visual-language data. Separate SFT and reward-optimized planners support different planning modes and evaluation settings.
The release is useful for autonomous-driving research involving interpretable perception and language-conditioned planning. ModelScope provides weights and configurations, with code and demo data in the companion repository. It is a research model requiring an appropriate evaluation and integration environment, not an independently validated vehicle-driving product.
