VLA-0: Building State-of-the-Art VLAs with Zero Modification
About
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 $\pi_0.5$-KI, OpenVLA-OFT and SmolVLA. Furthermore, without large-scale robotics-specific training, it outperforms methods trained on large-scale robotic data, like $\pi_0.5$-KI, $\pi_0$, 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/.
Key insights
How does VLA-0 turn a VLM into a robot action model?
VLA-0 represents continuous robot actions as space-separated integer text strings, using the VLM's native text generation with no new tokens, vocabulary changes, or architectural additions. A recipe of action ensembling and masked action augmentation unlocks performance.
What separates VLA-0 from prior VLA designs?
Unlike discrete-token, generative-head, or custom-architecture VLAs, VLA-0 needs no VLM modification. On LIBERO it averages 94.7 success, beating non-pretrained π0.5-KI (93.3) and pretrained GR00T-N1 (93.9), ranking second overall.
Where is VLA-0 demonstrated useful?
VLA-0 is demonstrated as a research prototype on the LIBERO simulation benchmark and a real SO-100 robot. On four real tasks it averages 60% success, outperforming SmolVLA by 12.5 points despite lacking large-scale pretraining.
What limitations do the authors acknowledge?
The authors state two explicit limitations: VLA-0 has not been evaluated with large-scale action pretraining, and its 4 Hz inference speed on a 5090 GPU could be improved via quantization or distillation, which they leave as future work.