# Metatron v1: NVIDIA's Unified Humanoid Control Policy for the Unitree G1 — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/f47c8a9190e41a77

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-30 07:09 ingested 2026-10-31 02:40

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xj8kzn/metatron_v1_nvidias_unified_humanoid_control/)

## retained text



```
Metatron v1 is a single learned policy that handles walking, scene interaction and fall recovery for the 29-joint Unitree G1 humanoid. It's a MaskedMimic-style transformer student, released under the NVIDIA Open Model License and integrated with ProtoMotions. The release comes in three non-interchangeable packages.

1. a native PyTorch checkpoint (about 459 MB) that keeps the posterior and normalization stats and supports simulator and supervised real-robot workflows.  
2. a single-robot browser ONNX pipeline, demo/model.onnx, with 39 inputs that outputs joint targets, PD gains and raw actions.  
3. a six-robot batched ONNX network, racer/network.onnx, that relies on external observation processing.

All three share the same epoch-6410 weights, and both ONNX exports run on MuJoCo WASM and ONNX Runtime Web via WebGPU or WASM.
```



document fetched 2026-10-31 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xj8kzn/metatron_v1_nvidias_unified_humanoid_control
