SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data