SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control
Paper โข 2511.07820 โข Published
SONIC (Supersizing Motion Tracking) โ NVIDIA's humanoid behavior foundation model for natural whole-body control. Provides a unified policy for walking, running, jumping, manipulation, and more.
| Field | Value |
|---|---|
| License | NVIDIA Open Model License (commercial use โ ) |
| Robot | Unitree G1 (29-DoF) |
| Input | Observations (joint states, IMU, motion references) |
| Output | 29-dim joint position targets |
| Control Frequency | 50Hz |
| Format | ONNX (encoder + decoder + planner) |
| Variant | Use Case | Lookahead |
|---|---|---|
| default | General-purpose motion tracking | 200ms |
| low_latency | Teleoperation, VLA execution | 80ms |
| sonic_v1_1 | Heading-normalized teleoperation | 200ms |
from huggingface_hub import snapshot_download
import onnxruntime as ort
import numpy as np
# Download model
snapshot_download(repo_id="BlackCatRoboticsAI/g1-sonic-base", local_dir="./g1-sonic")
# Load models
encoder = ort.InferenceSession("./g1-sonic/model_encoder.onnx")
decoder = ort.InferenceSession("./g1-sonic/model_decoder.onnx")
# Run inference
obs = np.random.randn(1, 1762).astype(np.float32)
tokens = encoder.run(None, {"obs_dict": obs})
action = decoder.run(None, {"obs_dict": obs})
SONIC supports real-time whole-body teleoperation via PICO VR headset:
@article{luo2025sonic,
title={SONIC: Supersizing Motion Tracking for Natural Humanoid Whole-Body Control},
author={Luo, Zhengyi and Yuan, Ye and Wang, Tingwu and others},
journal={arXiv preprint arXiv:2511.07820},
year={2025}
}
NVIDIA Open Model License โ commercial use permitted with attribution. See NVIDIA Open Model License for full terms.