Tasks and inputs
Metatron-v1 is listed for Data & control, Reinforcement learning. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Metatron-v1 is NVIDIA's open unified humanoid policy for the Unitree G1 that walks, interacts with scenes and recovers from falls, even in a browser.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights for self-hosting; no hosted API price. Hardware cost is yours.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
One policy for walking, interacting with obstacles and getting back up. That is the pitch for Metatron-v1, NVIDIA's humanoid controller for the Unitree G1. There is no separate recovery module to switch to after a stumble; the same learned network handles all three behaviours across the robot's 29 controlled joints.
It is also unusually easy to see running. Two of the artifacts are ONNX exports built for the browser.
| Package | What it is | Size |
|---|---|---|
G1/ |
Native PyTorch checkpoint, resolved config and reconstruction recipe | 459,455,983 bytes |
demo/model.onnx |
Single-robot browser pipeline including observation processing and action conversion | ~403 MB |
racer/network.onnx |
Six-robot batched network for G1 Racer; needs separate observation processing | ~403 MB |
The native checkpoint keeps the posterior and normalisation statistics. The browser graphs export only the deterministic prior and decoder path. A machine-readable config.json lists sizes and SHA-256 digests for the downloader.
A MaskedMimic-style transformer student with a deterministic learned prior and decoder, captured at epoch 6410 and capable of SAMP-style behaviour. Browser inference runs in FP32 on ONNX Runtime Web with WebGPU or WebAssembly. The surrounding application supplies physics, observations, goals and control logic; the ONNX files alone are not a demo.
Integration is with ProtoMotions. No public ProtoMotions release tag is asserted, so pin a Hub revision when you download. The weights repository went live on 27 October 2026.
The native G1 package supports simulator workflows and supervised deployment on a real G1. The browser exports are a different matter: matching their interface is not a safety certification, and those exact files have not been validated on hardware.
Start in simulation. Push the robot from several directions at increasing force and log how often it recovers, then add terrain and obstacle courses from your own scenes. Compare that recovery rate to whatever controller you run now. Only then move to a tethered real-robot test.
Licensing is the NVIDIA Open Model License for weights and configs, Apache-2.0 for code; training motion clips are not redistributed.
For motion generation that could feed a controller like this, see MotionBricks and other entries in the reinforcement learning directory. NVIDIA's robotics catalogue is on the provider page.
Compare the model on the work you need to complete.
Metatron-v1 is listed for Data & control, Reinforcement learning. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under NVIDIA Open Model License (weights); Apache-2.0 (code). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
runway
Praxis-1 is Runway's open 27B world action model for robotics, predicting video and robot actions together; Apache 2.0, LeRobot and vLLM-Omni support.
sensetime
SenseNova-RoboRSI-27B-Preview is SenseTime's open 27B vision-language model for embodied agents, post-trained from Qwen3.8-27B by recursive self-improvement.
OpenBMB
MiniCPM-RobotManip is OpenBMB's open 1.5B vision-language-action model for robot manipulation, keeping a minute of visual memory on-device.
OpenBMB
MiniCPM-RobotTrack is OpenBMB's compact vision-language-action policy on MiniCPM4-0.5B that follows a described person on a Unitree Go2 at 5+ FPS on board.
DHDRL
Hugging Face repository for Reinforcement learning.
DocPereira
DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.