DynaFLIP visual backbone outperforms DINOv2 and SigLIP by 22.5% on real-world robot manipulation tasks · Digg