Motion-DynVLA

Motion-DynVLA is a motion-dynamics extension of NVIDIA GR00T N1.6 for few-shot embodied adaptation. It retains the pretrained vision-language-action backbone and action expert, then adds midpoint motion identification and a reusable bank of 1024 motion prototypes.

Checkpoints

Folder	Description
libero-source	LIBERO source-domain checkpoint after dynamics warmup and joint training
libero-target20	LIBERO checkpoint adapted with 20 demonstrations per target task

Each folder is directly loadable with the Motion-DynVLA code through AutoModel.from_pretrained and AutoProcessor.from_pretrained.

Validated results

Setting	Result
LIBERO with 20 demonstrations per task	85.6 percent success over 500 fixed episodes
LIBERO gain from the pretrained bank over random initialization	3.6 percentage points

The public checkpoints contain inference weights and processor assets. Optimizer, scheduler, RNG and trainer state files are excluded.

Loading

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="trantor2nd/Motion-DynVLA",
    allow_patterns=["libero-target20/*"],
    local_dir="./checkpoints/Motion-DynVLA",
)

Code

The implementation, fixed manifests and evaluation tools are available in Motion-DynVLA.

Base model and benchmarks

Motion-DynVLA builds on NVIDIA GR00T N1.6. The published checkpoints are validated with paired evaluation on LIBERO.

License

These checkpoints are released under the NVIDIA License included with the code repository. Use is limited to non-commercial research under those terms. Third-party components remain subject to their respective licenses.