SONIC Motion Tracking

Humanoid MoCap API GPU Required

Overview

NVIDIA GEAR-SONIC locomotion controller with two operational modes:

Planner mode — velocity commands → full-body pose trajectories (10 Hz planning, 50 Hz output)
Tracking mode — reproduce motion capture clips via encoder+decoder latent-space pipeline

Architecture: Encoder + Decoder Pipeline

MoCap Clip

CSV joint angles

Encoder

model_encoder.onnx

Latent Space

z ∈ R^d

Decoder

model_decoder.onnx

Joint Targets

29-DOF output

ONNX Models

Model	File	Purpose
Planner	planner_sonic.onnx	Velocity → full-body trajectory
Encoder	model_encoder.onnx	MoCap clip → latent code
Decoder	model_decoder.onnx	Latent code → joint targets

Models are downloaded from nvidia/GEAR-SONIC on HuggingFace on first use.

Usage: Planner Mode

from roboharness.robots.unitree_g1 import (
    SonicLocomotionController, SonicMode
)

ctrl = SonicLocomotionController()
action = ctrl.compute(
    command={"velocity": [0.3, 0.0, 0.0], "mode": SonicMode.WALK},
    state={"qpos": qpos, "qvel": qvel},
)

Usage: Tracking Mode

from roboharness.robots.unitree_g1 import (
    SonicLocomotionController, MotionClip
)

ctrl = SonicLocomotionController()
clip = MotionClip.from_csv_dir("path/to/dance_clip/")
ctrl.set_tracking_clip(clip)

action = ctrl.compute(
    command={"tracking": True},
    state={"qpos": qpos, "qvel": qvel},
)

Note: This demo requires an NVIDIA GPU with CUDA/TensorRT and cannot run in CI. The page documents the API and architecture. See #86 (Phase 1) and #92 (Phase 2).

Generated by Roboharness