September 6, AriQuestionsthis weekOpen SourceQwen-Drive-1.0-4B, an autopilot based on Qwen3.5-4Bvisual language model.
Officially, Qwen-Drive-1.0First visual language base model for autopilotAT THE PRE-TRAINING STAGE, 3D PERCEPTION AND VISUAL QUESTIONS AND ANSWERS WERE HARMONIZED AND FURTHER EXTENDED TO SPORTS PLANNING, WHILE THE ORIGINAL VLM STRUCTURE FOR PRE-TRAINING WAS FULLY MAINTAINED。
The model uses a phased training programme that combines driving supervision with universal visual language data and retains universal visual understanding and command compliance while acquiring specific driving capabilities. Qwen-Drive-1.0-4B also provides two planning specialists, SFT and RL, with official assessments covering 3D perception, driving scenario understanding, universal visual language skills, and open, pseudo-closure and closed circle movement planning。
Description, Qwen-Drive-1.0 Only open-source construction of planned samplesTHE TRAJECTORIES OF THE VARIOUS DATA SETS WERE HARMONIZED AND A COMPREHENSIVE ASSESSMENT WAS CONDUCTED FROM THE OPEN-RING PREDICTION TO CLOSED-RING DRIVING. THE SFT MODEL IS COMPETITIVE ON ALL BENCHMARKS. AFTER INTENSIVE LEARNING, THE MODEL IS ONLY AT THE EXPENSE OF SMALL OPEN-RING ERROR, IN EXCHANGE FOR A COMPREHENSIVE UPGRADING IN HUMAN PREFERENCES AND THE SAFETY OF CLOSED-RINGS。
1AI WITH RELEVANT LINKS:
- GitHub: https://github.com/QwenLM/Qwen-Drive-1.0
- Hugging Face: https://huggingface.co/Qwen/Qwen-Drive-1.0-4B
- ModelScope: https://modelsscope.cn/models/Qwen/Qwen-Drive-1.0-4B