VLA Gallery / All models / Humanoid whole-body controllers

SONIC

SONIC (Supersizing Motion Tracking for Natural Humanoid Whole-Body Control)

NVIDIA · 2025-11

VAHumanoid whole-body controllers1.2M~42MRL policy

A general token-based whole-body controller that scales motion tracking in data·model·compute.

Paper

Backbone LLM: None
Vision encoder: —
Action head: MLP control decoder (RL policy)
Chunk: Not disclosed
Control rate: 50 Hz
Architecture (original summary): Quantizes VR·video·text·music·VLA inputs into FSQ universal tokens (64-dim latent motion tokens); MLP encoder (v1 [2048, 1024, 512, 512]) + control decoder (v1 [2048, 2048, 1024, 1024, 512, 512]; 2025-12-8 version [4096, 4096, 2048, 2048, 1024, 1024, 512, 512]) → 29-dim diagonal Gaussian actions, 50 Hz

Pretraining accelerators: 128 GPUs (model not disclosed); latest version scaling experiments with 16 · 32 · 128 GPUs
Pretraining duration: v1: 3 days on 128 GPUs; v1 'converges in 3~7 days' (unclear whether controller or planner); latest version '7 days'†
Pretraining steps & batch: 50k iterations · 4,096 parallel env per GPU (Isaac Lab)
GPU/TPU-hours: 21,000 (Stated · GPU-hours (GPU model Not disclosed); latest version (v4) 21k, v1 abstract 9k·v1 §2.1 32,000 (conflict across versions))
Fine-tuning: N/A (controller); GR00T N1.5 fine-tuned to output SONIC tokens, GPU not disclosed
Data type: Motion capture
Data scale: v1: in-house 700 hours·100M+ frames (50 Hz)·170 people; latest version after retargeting 611 hours·317,189 training clips·8,447 subcategories
Main datasets: In-house motion capture; public subset BONES-SEED (142K+ motions, about 288 hours, released 2026-03-16)
Tasks & evaluation: 42M test-content success rate 99.6% · MPJPE-L 23.8 mm (1.2M 98.0% · 27.7 mm); real-world 124 sequences 99.2%†; loco-manipulation with GR00T N1.5 average 75% over 5 tasks
Robot: Unitree G1

Inference device & speed: Policy 50 Hz, kinematic planner 10 Hz, user input 100 Hz, motors 500 Hz; planner latency under 5 ms on a laptop, 12 ms on Jetson Orin; deployed with C++ and TensorRT
Inference memory: Not disclosed
Weights bf16 (calc.): 0.1 GB
Where it runs: Onboard (Jetson Orin, planner 12 ms)

Reinforcement learning: Used (RL controller training) · Large-scale RL motion tracking in Isaac Lab (128 GPUs, 4,096 envs per GPU)
Base control: Humanoid legged (RL policy drives joints directly) · General motion tokens → control decoder directly outputs 29-dim joint targets for locomotion (velocity commands 0–6 m/s); when paired with a VLA, the VLA outputs navigation commands

Key features: A general token-based whole-body controller that scales motion tracking in data·model·compute. Scaling an MLP (instead of a transformer) from 1.2M to 42M raised test-content success from 98.0% to 99.6%.
Notes: Data hours are the v1 in-house dataset of 700 hours (611 hours after retargeting in the latest version). GPU-hours conflict across versions: v1 abstract 9k vs v1 §2.1 32,000 ('128 GPUs over 3 days'); the latest version trains 16·32·128 GPUs all to 50k iterations, giving about 2k·9k·21k. VLA data 3,900 trajectories (13 objects × 300)·multi-object 1,000 trajectories†; v1-line PDF fine-tunes GR00T N1.5 with 300 teleoperation demonstrations. No weights at 2025 release; checkpoints 2026-02-19·training code 2026-04-10 (weights NVIDIA Open Model License, code Apache 2.0). The 2026 training guide recommends 64+ GPUs (not a 2025 paper value). v4 2026-08-13 (Science Robotics 11(117) eaed4592). Kinematic planner is 'Transformers or Conv1D'.
Verification: Confirmed