NEO-VLA

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

Paper · 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
stage2_backbone/	Vision-language backbone cropped from stage2/	75,000
stage_libero/	LIBERO fine-tuned policy	—
stage_rtw/	RoboTwin 2.0 fine-tuned policy	—

Stage 2 initializes from stage 1 at step 113,400. stage_libero/ and stage_rtw/ are policies fine-tuned for the LIBERO and RoboTwin 2.0 simulation benchmarks, released for inference; their files do not record a training step.

stage2_backbone/ is the vision-language backbone taken from the stage-2 checkpoint at step 75,000 and saved as a standalone Hugging Face model, with the action expert removed. It is not the original NEO-2B (NEOpp-2B-SFT) checkpoint, which is not released. The backbone was fully fine-tuned during stage-2 joint vision-language and action training, so essentially every tensor differs from the original. EMA weights are used where available. It is a post-training artifact of NEO-VLA, not a general-purpose replacement for NEO-2B.

Multimodal understanding performance of stage2_backbone:

Benchmark	Score
RealWorldQA	66.80
AI2D (test)	82.87
ChartQA (test)	80.76
TextVQA (val)	76.59
MMBench-EN (dev)	82.65
MMBench-EN (dev, circular)	74.48

Files

Each stage*/ checkpoint directory contains the native state.pt checkpoint, meta.json and its own embodiments.json normalization records. The stage1/ and stage2/ checkpoints include model weights and native training state (optimizer, scheduler and EMA). stage_libero/ contains model weights only; stage_rtw/ contains model and EMA weights. Use the normalization records supplied with that checkpoint; the action representation and masks depend on them. stage2_backbone/ instead holds model.safetensors with its configuration, tokenizer and modeling code, and no meta.json or embodiments.json. Use the original NEOpp-2B-SFT tokenizer and the policy configuration supplied with the code.

Download checkpoints with huggingface_hub; keep only the directories you need in allow_patterns:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="weepiess2383/NEO-VLA",
    allow_patterns=["stage1/*", "stage2/*", "stage2_backbone/*", "stage_libero/*", "stage_rtw/*"],
    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. The stage*/ checkpoints (except stage2_backbone/) 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.