VLA-0: Building State-of-the-Art VLAs with Zero Modification

TL;DR

This work investigates how to build Vision-Language-Action models without modifying the underlying vision-language model. By prompting a VLM to output actions as text and applying a careful training/testing recipe (including masked action augmentation and action ensembling), VLA-0 achieves competitive or state-of-the-art performance on LIBERO and real-world tasks without large-scale robotics pretraining. The findings suggest that simple, text-based action decoding can rival more complex approaches, with substantial implications for efficiency and generalization in robotic manipulation. The approach is validated through comprehensive simulations, real-world experiments, and targeted ablations, underscoring the practical potential of zero-modification VLAs.

Abstract

Vision-Language-Action models (VLAs) hold immense promise for enabling generalist robot manipulation. However, the best way to build them remains an open question. Current approaches often add complexity, such as modifying the existing vocabulary of a Vision-Language Model (VLM) with action tokens or introducing special action heads. Curiously, the simplest strategy of representing actions directly as text has remained largely unexplored. This work introduces VLA-0 to investigate this idea. We find that VLA-0 is not only effective; it is surprisingly powerful. With the right design, VLA-0 outperforms more involved models. On LIBERO, a popular benchmark for evaluating VLAs, VLA-0 outperforms all existing methods trained on the same robotic data, including $π_0.5$-KI, OpenVLA-OFT and SmolVLA. Furthermore, without large-scale robotics-specific training, it outperforms methods trained on large-scale robotic data, like $π_0.5$-KI, $π_0$, GR00T-N1 and MolmoAct. These findings also translate to the real world, where VLA-0 outperforms SmolVLA, a VLA model pre-trained on large-scale real data. This paper summarizes our unexpected findings and spells out the specific techniques required to unlock the high performance of this simple yet potent VLA design. Visual results, code, and trained models are provided here: https://vla0.github.io/.

Figures (4)

• Figure 1: Schematic representation of VLA-0. VLA-0 converts a VLM into a VLA by prompting the VLM to predict action as text. This strategy is surprisingly effective and achieves state-of-the-art results akin to alternatives.
• Figure 2: Families of methods for building VLAs. We categorize existing VLAs into three categories: Discrete Token VLAs, Generative Action Head VLAs and Custom Architecture VLAs. In this work, we propose the VLA-0 family where the VLM is prompted to directly predict action as text. Unlike other methods, VLA-0 requires no change to the underlying VLM.
• Figure 3: Our proposed VLA-0. It creates a VLA without making any changes to the underlying VLM. It takes a system prompt, language instruction, and images as input, and outputs actions represented as space-separated integers.
• Figure 4: Performance on Real. We compare VLA-0 to SmolVLA on four different tasks with SO-100. VLA-0 outpeforms SmolVLA on average. SmolVLA is pretrained with large-scale SO-100 data while VLA-0 is trained from scratch.