SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

SmolVLA   2025   2506.01844  
arXiv Code

作者们的motivation就是去造一个efficient的VLA,本质上完成的内容还是

  1. 一个轻量化推理的VLM,且只用其中的一半的: “In practice, we find setting N to half the total layers (N = L/2) offers a good tradeoff between speed and performance, effectively halving the LLM and action expert’s computational cost.”
  2. self-attention和cross-attention交替进行的flow-matching
  3. async inference,感觉可以用RTC取代掉 但是有一点值得学习的是他们对于async action buffer的可视化,感觉之后可以用在我们的研究上。


← Home · All Readings