Early Fusion Helps Vision Language Action Models Generalize Better

Huang Huang, Fangchen Liu, Letian Fu, Tingfan Wu, Mustafa Mukadam, Jitendra Malik, Ken Goldberg, Pieter Abbeel

Keywords: vision language action model; robot foundation model

Abstract: Recent advances in Vision-Language-Action (VLA) models can enable robots to perform a wide range of tasks based on language or goal-based instructions. These VLA models typically encode text and images into disjoint tokens, generating actions that align with the given instructions. This requires the VLA models to simultaneously perform vision-language understanding and precise closed-loop control, resulting in significant challenges for them to generalize to new environments. However, contrastive pre-trained VLMs, such as CLIP, already possess vision-language alignment capabilities, which are underutilized by current VLA models. In this paper, we propose Early Fusion VLA (EF-VLA), a novel VLA architecture that exploits CLIP’s vision-language understanding by performing early fusion, extracting fine-grained vision-language tokens relevant to the task instructions before passing them to the transformer policy. EF-VLA keeps the VLM frozen, allowing it to effectively perform unseen tasks without requiring fine-tuning, which often reduces generalization capabilities. Simulation and real-world experiments suggest that EF-VLA outperforms state-of-the-art VLA models on diverse tasks, with significant generalization capabilities in unseen environments.

Previous Publication: No

Submission Number: 10