VLAICML 20252025
OpenVLA-2: Open-Source Vision-Language-Action Model
Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Chelsea FinnStanford University, UC Berkeley
摘要
We present OpenVLA-2, an open-source vision-language-action model for general-purpose robot control. OpenVLA-2 combines a powerful vision-language backbone with an action tokenizer, enabling robots to follow natural language instructions for diverse manipulation tasks with strong zero-shot generalization.
open-source VLArobot controlvision-languageaction tokenizerzero-shot generalization
技术细节
模型骨架
SigLIP-So400m + Llama 2 7B
编码器
SigLIP So400m/14 + Proprioceptive MLP
解码器
Llama 2 7B Action Decoder (32 layers)
