Dino-Flex-Qwen aka DFQ VLA

DFQ is a ~1B parameter VLA that processes multi-camera, multi-timestamp images and acts as a policy for autonomous driving. It is built using NVIDIA's PhysicalAI-Autonomous-Vehicles dataset

Model Architecture

Key components of DFQ VLA are:

• DinoV3 as vision encoder: Processes 16 images to output roughly 10k tokens
• Flex scene encoder encoder vision tokens into 900 scene tokens using joint self attention
• Qwen3-06.B LLM consumes vision tokens + trajectory history to produce 8 Meta actions
• Action chunking head consumes these meta actions + last hidden state of LLM to produce refined 64 xyz + 3x3 rotations

Current Status and ToDos:

• SFT: in progress
• Behaviour tuning: ToDo
• Integration with AlpaSim: ToDo
• Checkpoints available on Hub: ToDo