NVIDIA Releases Metatron-v1 Reinforcement Learning Model

NVIDIA has released Metatron-v1, a reinforcement learning model designed for interactive humanoid control. The model is hosted in a repository and integrated with ProtoMotions. According to the model documentation on Hugging Face, the model is intended for researchers and developers running interactive humanoid control in a browser, though separate native G1 packages also support simulator and supervised real-G1 deployment workflows.

Model Architecture and Technical Specifications

The Metatron v1 model is a "MaskedMimic-style transformer student with deterministic prior and decoder." It is designed for the Unitree G1 robot, which features 29 controlled joints. The model uses FP32 inference and a deterministic learned prior.

NVIDIA provides several artifacts for the model:

• Native PyTorch G1 inference: This includes a checkpoint, resolved config, reconstruction recipe, and legal files. The native checkpoint is 459,455,983 bytes.
• demo/model.onnx: A single-robot browser scene/recovery pipeline. This file is 403,375,789 bytes and includes observation processing and action conversion.
• racer/network.onnx: A six-robot batched G1 Racer network. This file is 403,036,698 bytes and requires the G1 Racer's separate observation processing.

NVIDIA notes that the ONNX inference exports are for a unified Metatron G1 scene/recovery policy. This policy covers walking, scene interaction, and recovery, and does not use a separate recovery adapter. The two ONNX files use the same learned policy weights but have different runtime interfaces and are not interchangeable.

Deployment and Integration

The browser integrations utilize MuJoCo WASM for simulation and ONNX Runtime Web for inference. While WebGPU is recommended for capable devices, CPU/WASM serves as a fallback, though it does not guarantee real-time operation for six robots.

The demo pipeline consumes 39 inputs, including the current G1 state, eight historical state samples, paired incoming raw actions, a nine-entry raw-action history, masked future pose/style/velocity goals, and scene-box context. The model returns joint PD targets (action), PD gains (stiffness, damping), and raw_action for history feedback.

Limitations and Safety

NVIDIA includes several warnings regarding the use of the model. The company states that "Recovery, balance and tracking can fail, especially outside the evaluated states, terrain and command distribution." Furthermore, the model is not a collision-avoidance system or a safety controller.

Regarding hardware use, NVIDIA specifies that it "requires independent validation, appropriate joint/torque limits, supervised testing, and an emergency-stop mechanism." The company also notes that certain features, such as game launch assists, low-gravity zones, and respawns, are application mechanics rather than capabilities guaranteed by the learned policy or the physical robot.

Why it matters

Metatron-v1 provides a framework for testing humanoid locomotion and recovery. While the ONNX files provide the learned policy, they are not complete demos; the applications provide the necessary physics, observations, task goals, and control logic.

---

Sources

• huggingface.co: nvidia/Metatron-v1 · Hugging Face

This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.