What Is VLA-0?

VLA-0 is introduced in VLA-0: Building State-of-the-Art VLAs with Zero Modification (arXiv 2510.13054) by Ankit Goyal, Hugo Hadfield, Xuning Yang, Valts Blukis and Fabio Ramos of NVIDIA, with an official project page and GitHub repository (NVlabs/vla0). Read: the arXiv abstract page and full paper text, the project page and the GitHub README. Check the version you cite.

VLA-0 is third-party research. Firsthand is not part of it or affiliated with its authors, and makes no claim that any model was trained on Firsthand data. Firsthand captures human egocentric video, not robot data. What the site states is custom collection of real human demonstrations, fast and on demand, to a written spec: 3D hand pose at 21 joints per hand, calibrated stereo depth and 6-DoF IMUs at 200 Hz, delivered in RLDS, LeRobot format or WebDataset with a buyer-owned commercial license. The only language data the site states is one task instruction per episode. Whether that fits a VLA-0 style prompt is Not stated.

Per the abstract, the best way to build vision-language-action models remains an open question. Current approaches often add complexity, such as modifying a vision-language model’s vocabulary with action tokens or introducing special action heads.

The authors write that the simplest strategy, representing actions directly as text, has remained largely unexplored, and introduce VLA-0 to test it.

Per the paper, VLA-0 keeps the model intact: it introduces no new tokens, alters no vocabulary and adds no layers. The input is a system prompt, one or more images and a task instruction, and the output is a sequence of space-separated integers.

Continuous action values are normalized to a fixed integer range, with 0 to 1000 given as an example, so resolution is set by that range rather than by the vocabulary. The authors build on a 3-billion-parameter Qwen-VL-2.5 model and say the method applies to any other vision-language model.

The authors name three components: action decoding, ensemble prediction and masked action augmentation. At each step the model predicts several future actions, and the predictions available for the current step are averaged, a technique they credit to the Action-Chunking Transformer.

During training, characters in the target action string are randomly masked, so the model has to rely on the images and instruction rather than auto-completing numbers. Training is full fine-tuning with cross-entropy loss: 64 epochs, batch size 192, learning rate 5e-6, about 32 hours on 8 A100 GPUs, per the authors.

On LIBERO, the authors report a 94.7 average success rate, the best among models without large-scale robot-action pretraining. Against pretrained models they report an average rank of 2.8, behind OpenVLA-OFT at 1.5.

In the real world, on an SO-100 robot with LeRobot, four tasks and 100 demonstrations per task, they report VLA-0 outperforming SmolVLA by 12.5 points, at 4 Hz inference on a desktop with a 5090 GPU. These are the authors’ results. Firsthand does not reproduce them and ranks no models.

On LIBERO, removing ensembling lowered the average success rate by 2.0 points and removing masked action augmentation by 1.2. An action resolution of 250 lowered it by 1.5, while 4000 gave no further gain over 1000. Tiling images into one composite versus feeding them separately made no discernible difference.

The paper groups OpenVLA with discrete-token VLAs, which map action bins to vocabulary tokens, and π-FAST with custom designs. The site’s FAST entry describes a discrete cosine transform, quantization and byte-pair encoding that turn action chunks into tokens. VLA-0 instead writes the numbers out as text.

The site’s VLA-Adapter entry covers a lightweight Policy module with Bridge Attention. How VLA-0 relates to it is Not stated in the sources read.

The authors list two open areas: how VLA-0 performs when trained with large-scale action data, and faster inference through quantization and distillation. Their example instruction is “put the banana on the plate”. On the site, the only language data is one task instruction per episode.

The GitHub README states that the code and the LIBERO model are released under CC BY-NC 4.0 and that the base model is subject to the Qwen Research License. Check both before any use.

When a program needs real human egocentric demonstrations that public data does not cover, specified in writing. Firsthand does not build or train models. Confirm the spec on a Spec Pilot, a small fast first batch; the median from signed spec to first validated batch is 48 hours where coverage is deep, three to four weeks where thin. Billing is per validated hour.

What is VLA-0?

A published vision-language-action design (arXiv 2510.13054) from NVIDIA authors in which a pretrained vision-language model is prompted to output robot actions as text. It is third-party research.

What does zero modification mean?

Per the paper, VLA-0 introduces no new tokens, alters no vocabulary and adds no new layers to the vision-language model. Actions are written as space-separated integers using the model’s native text generation.

What did the authors report on LIBERO?

A 94.7 average success rate, the best among models without large-scale robot-action pretraining, and an average rank of 2.8 against pretrained models. These are the authors’ figures.

Is the VLA-0 code available?

Yes. The authors link an official GitHub repository (NVlabs/vla0) and project page. The README states CC BY-NC 4.0 for the code and LIBERO model, with the base model under the Qwen Research License.

Does Firsthand data train VLA-0?

No claim is made. Firsthand is not affiliated with its authors, captures human egocentric video rather than robot data, and does not state that any model was trained on Firsthand data. The only language data the site states is one task instruction per episode.