VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model trained on CALVIN ABC-D.

We have developed and released the VLA-Adapter family of VLA models, a series of fine-tuned generative action models. The VLA-Adapter VLM follows the Prismatic-VLM architecture, using only a very small backbone (Qwen2.5-0.5B) for the LLM. On common robotics benchmarks, it surpasses open-source VLA models with 8.5B, 7B, 4B, 3B, and 2B backbones.

Input: Models input image and text.

Output: Models generate action only.

Model Architecture: The VLA-Adapter consists of a VLM for receiving and processing image and text information and a policy for generating actions. We systematically analyzed the benefits that the VLM provides to different types of policy conditions and determined a unified framework. We then utilized our designed Bridge Attention module to fuse the conditions generated by the VLM with the initial action information in the policy, bridging the gap between VL and A to the greatest extent possible. This resulted in a high-performance VLA model on a tiny-scale backbone.