Selector guide · from the official model card

Which SONIC model do you need? Three checkpoints, one decision

GEAR-SONIC ships three whole-body controller checkpoints for the Unitree G1. They differ in reference horizon and deployment intent — pick by how you will use the controller, not by version number. Everything below is transcribed from the official model card.

Available models

ModelSMPL reference inputIntended use
Default SONIC (original release) 10 future frames at 20 ms spacing, approximately 200 ms of reference lookahead Default general-purpose SONIC controller for motion tracking, planning, teleoperation, and compatibility with existing deployments. G1 and teleoperation future-reference observations use step5.
Low-latency teleoperation 4 future frames at 20 ms spacing, approximately 80 ms of reference lookahead Intended for more responsive whole-body teleoperation and VLA execution. G1 and teleoperation future-reference observations use step1. Use its encoder, decoder, and observation config together. HF files →
SONIC v1.1 10 future frames at 20 ms spacing, approximately 200 ms of reference lookahead Uses robot-heading-normalized target orientation and was trained with wrist-pose augmentation. Intended for heading-stable whole-body teleoperation and SONIC-backed VLA policies that use this controller. G1 and teleoperation future-reference observations use step5; this is not the low-latency model. HF files →
All three share the same controller core: All three models use the SONIC universal-token controller, produce 64-dimensional latent motion tokens, run the controller at 50 Hz, and support SMPL pose, G1 motion reference, and teleoperation inputs.
Read before benchmarking: The lookahead values describe the reference horizon presented to the controller. They are not measurements of total end-to-end teleoperation latency, which also includes sensing, networking, preprocessing, and inference.

Pick by use case

Default SONIC

Motion tracking, planning, teleoperation, and compatibility with existing deployments. The safe default when you are not sure — it is the original general-purpose release.

Low-latency teleoperation

Most responsive whole-body teleoperation and VLA execution. 4-frame lookahead (~80 ms). Use its encoder, decoder, and observation config together — do not mix with the default files.

SONIC v1.1

Heading-stable whole-body teleoperation and SONIC-backed VLA policies. Robot-heading-normalized target orientation, trained with wrist-pose augmentation. Not the low-latency model.

Released files

ModelDeployment filesPyTorch + config
Default SONICmodel_encoder.onnx, model_decoder.onnx, observation_config.yamlsonic_release/last.pt, sonic_release/config.yaml
Low-latency teleoperationlow_latency/model_encoder.onnx, low_latency/model_decoder.onnx, low_latency/observation_config.yamllow_latency/last.pt, low_latency/config.yaml, low_latency/model_config.yaml
SONIC v1.1sonic_v1_1/model_encoder.onnx, sonic_v1_1/model_decoder.onnx, sonic_v1_1/observation_config.yamlsonic_v1_1/last.pt, sonic_v1_1/config.yaml, sonic_v1_1/model_config.yaml

Deployment uses C++ and TensorRT; the PyTorch checkpoints support Isaac Lab evaluation and continued training. Model weights are covered by the NVIDIA Open Model License.

SONIC model FAQ

Which SONIC model should I use?
Use Default SONIC for general tracking, planning, teleoperation and compatibility with existing deployments. Use Low-latency teleoperation for the most responsive whole-body teleoperation and VLA execution (4-frame, ~80 ms lookahead). Use SONIC v1.1 when you need heading-stable whole-body teleoperation or SONIC-backed VLA policies (10-frame, ~200 ms lookahead, robot-heading-normalized, wrist-pose augmented).
What is the difference between the SONIC checkpoints?
They differ in reference horizon and intended deployment: Default uses 10 future frames at 20 ms spacing (~200 ms lookahead, step5 observations), Low-latency uses 4 frames (~80 ms, step1), and v1.1 uses 10 frames (~200 ms) with robot-heading-normalized target orientation and wrist-pose augmentation.
Is lookahead the same as latency?
No. The lookahead values describe the reference horizon presented to the controller. They are not measurements of total end-to-end teleoperation latency, which also includes sensing, networking, preprocessing, and inference — this is stated verbatim in the official model card.
Where are the SONIC checkpoints hosted?
On Hugging Face under nvidia/GEAR-SONIC. Default SONIC files sit at the top level (model_encoder.onnx, model_decoder.onnx, observation_config.yaml, training checkpoint sonic_release/last.pt); low-latency under low_latency/ and v1.1 under sonic_v1_1/. Download with python download_from_hf.py in the repository.