SmolVLA

“SmolVLA [5] (~450M parameters from LeRobot), with the pretrained vision-language model frozen and the action expert fine-tuned”