Back home

Metatron-v1

hf+x+web 0.68
Visit source

Intro

Metatron-v1 is a humanoid robot control policy that NVIDIA published on Hugging Face as nvidia/Metatron-v1 late on 2026-10-27 (UTC). It is not a language model: it is a learned controller for the Unitree G1 humanoid that handles walking, scene interaction and fall recovery with a single policy.

The release is low-key — two ONNX files (~403 MB each), a license and a model card, with no NVIDIA press release or paper. A companion repo, nvidia/metatron-browser ("Metatron v1 G1 Browser Exports"), lets the policy run in a web page through ONNX Runtime Web and MuJoCo compiled to WebAssembly. It was surfaced on X by HuggingPapers as "the Metatron G1 humanoid walking and recovery policy".

Explain

Key points

  1. One unified policy: per the model card, the current export is an epoch-6410 "SAMP-capable unified student" that combines walking, scene interaction and recovery without a separate recovery adapter.
  2. Inputs: current G1 state plus a short history (eight past state samples), raw-action history, and masked future pose / style / velocity goals plus scene-box context (a MaskedMimic-style goal-inpainting setup).
  3. Outputs: joint PD targets (action), stiffness and damping gains, and raw_action for history feedback — i.e. control-rate motor commands, no text.
  4. Two export layouts, same weights: a single-robot demo graph with observation processing built in, and a six-robot batched "Racer" graph that expects observations computed outside the graph. They are not interchangeable.
  5. License: NVIDIA open model license.

Caveats

  • It is an export package, not a training kit; no hardware-validated benchmark is published, and demos are in simulation.
  • The name is one letter from NVIDIA's Nemotron LLM family but is unrelated to it.

Why this radar term

A new, distinctive NVIDIA model name that appeared on Hugging Face inside this scan window and was picked up on X the same evening.

Usage

How to try it

  1. Open nvidia/Metatron-v1 and read the model card for the exact tensor names and shapes.
  2. For the quickest look, use the nvidia/metatron-browser exports, which run the G1 policy in the browser with ONNX Runtime Web + MuJoCo WASM.
  3. In your own simulator, load the single-robot demo ONNX with ONNX Runtime, feed G1 state / history / goal tensors, and apply the returned PD targets and gains at control rate.
  4. Use the six-robot "Racer" export only if you compute observations outside the graph and batch multiple robots.

Good fit

  • Robotics researchers and hobbyists working with Unitree G1 in simulation
  • Teams studying unified locomotion + recovery policies and goal-conditioned control

Caveats

  • Validate carefully before any real-hardware deployment; NVIDIA has not published real-robot results for this release