Pre-trained Models

Each shipped checkpoint has a model card beside it with the training recipe, input and output contract, artifact roles, intended use, and limitations.

NVIDIA-provided pretrained models use the NVIDIA Open Model License unless their card explicitly specifies otherwise; ProtoMotions source code remains Apache-2.0. Each model directory includes a full license and attribution notice. Keep these files with redistributed checkpoints and exported models. See the model licensing policy for details. Training datasets and robot assets retain their separate terms.

Unless its model card says otherwise, a policy should be used on its training simulator. The recommended checkpoints were trained or fine-tuned in IsaacLab. The G1 deployment tracker is the only current policy trained with the full transfer-oriented domain-randomization recipe, so it is expected to transfer to compatible simulators and hardware; this is not a guarantee for every backend or version.

Model	Purpose	Runtime expectation	Model card
SOMA BONES-SEED GPC prior	Discrete GPC prior for SOMA motion	IsaacLab only	Read card
G1 BONES-SEED deployment tracker	Domain-randomized motion tracking and deployment	Trained in IsaacLab; expected to transfer	Read card
Metatron v1 G1 (MaskedMimic)	Motion, recovery and scene-conditioned control	Trained in IsaacLab; expected to transfer	Read card
SOMA BONES-SEED tracker	Continuous-action SOMA motion tracking	IsaacLab fine-tune only (inference_last_lab.ckpt)	Read card
SOMA BONES-SEED FSQ tracker	Discrete FSQ tracker used by GPC	IsaacLab only	Read card
Anatomical SOMA BONES-SEED FSQ tracker	Domain-randomized FSQ tracker for the anatomical SOMA humanoid	Trained in IsaacLab; verified in MuJoCo	Read card

Download Metatron v1 G1

The native package lives on Hugging Face, not in Git LFS. From the repository root, after installing ProtoMotions and the MuJoCo requirements:

python scripts/download_metatron.py
python protomotions/inference_agent.py --simulator mujoco --num-envs 1 \
    --checkpoint data/pretrained_models/metatron/G1/inference_last.ckpt \
    --inference-config examples/inference/metatron.py

The standard-library helper downloads about 460 MB, pins an immutable Hub revision, verifies sizes and SHA256 digests, and includes the model card, license, notice and all configuration sidecars. No account or token is required. Verified files are reused; an existing mismatched file is refused, not overwritten. Use --output-dir for another location and adjust --checkpoint accordingly. --configs-only fetches the small sidecars for configuration tests, without weights; it is not sufficient for inference. Keep the package files together.

The playback config starts the robot idle; pressing X in the viewer cycles four goal modes (idle, walk to a target location, follow a velocity command, reach with the right hand), and the scene stream is published empty by the saved configuration’s scene_boxes control. Its generic ONNX re-export takes the same eight scene_boxes_* inputs from the host, with a host contract that differs from the browser graphs’ (no simulator root pose; the current action and the eight-slot history as separate inputs). The separate demo/model.onnx and racer/network.onnx interfaces remain available in the same Hub repository; use each with its paired application. No pretrained Metatron SOMA policy is included.