NVIDIA Releases ONNX Inference Exports for Metatron G1 Scene/Recovery Policy

Metatron v1 — G1 Browser Exports

Overview

ONNX inference exports of the unified Metatron G1 scene/recovery policy used by the Metatron browser demo and G1 Racer. Both files use the same learned policy weights, with different runtime interfaces. They are not interchangeable. The current exports are the epoch-6410 SAMP-capable unified student. Walking, scene interaction and recovery share one policy, with no separate recovery adapter.

Field	Value
Robot	Unitree G1, 29 controlled joints
Model	MaskedMimic-style transformer student with deterministic prior and decoder
Inference	Deterministic learned prior; FP32
Format	ONNX
Integration	ONNX Runtime Web with WebGPU or WebAssembly

Intended Use

Interactive humanoid control, locomotion and recovery evaluation with the matching browser applications. The applications provide physics, observations, task goals and control logic; the ONNX files alone are not complete demos.

Metatron supports deployment workflows beyond the browser, including other simulators and real G1 robots. These particular browser exports require their matching observation and control interfaces: compatibility is not a safety certification or a claim that these exact files have been validated on hardware.

Artifacts

File	Interface	Size
`demo/model.onnx`	Single-robot pipeline, including observation processing and action conversion	403,375,789 bytes
`racer/network.onnx`	Six-robot batched network; requires G1 Racer's separate observation processing	403,036,698 bytes

SHA256 digests:

demo/model.onnx: 7180e82cae7ec04ff7766c262d430f57573d93678e025d8a71916885072a0cd4
racer/network.onnx: 130a22718a34c0da488e2157553b40efedfefcf6ca6cb6b97cfdd67f8dfbeaf3

The Racer export changes graph organization and batching, not learned weights. Neither artifact is a resumable training checkpoint. Runtime code, robot assets, motion clips and application configuration are distributed separately.

Inputs and Outputs

The demo pipeline consumes current G1 state, eight historical state samples, paired incoming raw actions, nine-entry raw-action history, masked future pose/style/velocity goals and scene-box context (39 inputs). There is no recovery_enabled input. It returns joint PD targets (action), PD gains (stiffness, damping) and raw_action for history feedback. Recovery targets remain live runtime inputs, not baked goals. Applications use staged local goals with independent per-robot recovery state.

G1 Racer computes observations outside the neural graph and evaluates all six robots in one batch. Its network exposes action and raw_action; controller parameters come from the paired runtime. Do not feed raw demo inputs directly to this network or change its batch size without rebuilding and validating it.

Preserve the paired application's joint/body ordering, quaternion convention, history ordering, normalization, target masks, control rate and action mapping. Use the matching runtime manifest and ONNX tensor signatures for exact layouts.

Runtime Support and Validation

The browser integrations use MuJoCo WASM for simulation and ONNX Runtime Web for inference. WebGPU is recommended on capable devices; CPU/WASM is a fallback and does not guarantee real-time six-robot operation. GPU drivers, browsers and hardware affect startup time and throughput.

The six-robot ONNX and compiled-observation paths were checked against the single-robot demo export on 60 distinct/history-perturbed test states, with maximum absolute action error 5.55e-6. Recovery height, recovery timing, action history and scene inputs were independently varied and affected the outputs. These are numerical export checks, not task-success or real-robot benchmarks. Racer's paired application also runs action-parity fixtures at startup.

For reproducibility, pin downloads to a repository commit rather than main. Public browser delivery requires anonymous access to the files; never embed a Hugging Face access token in client-side code.

Limitations and Safety

Recovery, balance and tracking can fail, especially outside the evaluated states, terrain and command distribution.
Game launch assists, low-gravity zones and respawns are application mechanics, not capabilities guaranteed by the learned policy or physical robot.
The model is not a collision-avoidance system or a safety controller.
Hardware use requires independent validation, appropriate joint/torque limits, supervised testing, and an emergency-stop mechanism.
These exports are not a complete training-reproduction package. No training dataset or training checkpoint is included.

License and Asset Rights

Model terms: NVIDIA Open Model License Agreement, with the required attribution in NOTICE.txt. The complete agreement governs; this card does not replace or extend it.

Model licensing does not grant rights to motion datasets, recovery reference clips, robot meshes, game assets or third-party runtime software. None of those assets is included in this model repository. Obtain them through their respective distributions and follow their separate terms. Permission to include a clip inside a demo or game does not imply permission to redistribute that clip as a standalone dataset.

These ONNX exports are publicly available for anonymous download. Their availability does not change the separate distribution terms or release status of the associated application bundles, motion clips and other assets.