NEO-VLA

Checkpoints for From Pixels to Actions: Towards Native Vision-Language-Action Policies.

Project page · Code · Dataset

Checkpoint	Training stage	Step
stage1/	Action pretraining with a frozen backbone	113,400
stage2/	Joint vision-language and action training	75,000

This stage-2 checkpoint initializes from stage 1 at step 113,400. Subsequent simulator midtraining and task-specific fine-tuning checkpoints are separate experiments.

Files

Each stage directory contains the native state.pt checkpoint, meta.json and its own embodiments.json normalization records. The checkpoint includes model weights and native training state. Use the normalization records supplied with that checkpoint; the action representation and masks depend on them. Use the original NEOpp-2B-SFT tokenizer and the policy configuration supplied with the code.

Download both stages with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="weepiess2383/NEO-VLA",
    allow_patterns=["stage1/*", "stage2/*"],
    local_dir="NEO-VLA",
)

For adaptation, set trainer.init_from to the downloaded NEO-VLA/stage2 directory, or NEO-VLA/stage1 to initialize from action pretraining. Architecture and training configuration are described in the training guide.

Use and limitations

These checkpoints support robot-learning research and require task-specific adaptation and evaluation before use on a physical robot. They are native NEO-VLA checkpoints, not Transformers AutoModel exports.

The original joint-training run used multimodal understanding co-training data. The released code provides COCO ShareGPT as a default configuration example; it does not reproduce that original data mixture by itself.

The code is distributed under Apache-2.0. Backbone weights and source datasets retain their upstream terms; the code license does not relicense those assets.