Hi, thanks for your great work!
I'm curious why you chose SigLIP instead of reusing Qwen2.5-VL's visual encoder as the input for low-level control. Could you share the reasoning behind this decision?