NVIDIA's SONIC: Scaling Laws for Natural Humanoid Whole-Body Control
yuewang314 · x · 2026-08-13
NVIDIA, in collaboration with USC and other institutions, has introduced SONIC, a foundation model for humanoid whole-body control, featured in Science Robotics.
The research highlights that while LLMs have successfully leveraged scaling laws, humanoid neural controllers remain limited in size, data, and compute. SONIC aims to build a generalist humanoid controller by scaling model capacity, data, and compute to achieve natural and robust whole-body movements.
Key Scaling Dimensions:
- Network Size: Scaled from 1.2M to 42M parameters.
- Dataset Volume: Trained on 100M+ frames of motion capture data (700 hours).
- Compute: Required 21,000 GPU hours.
SONIC positions motion tracking as a scalable task, leveraging dense supervision from diverse mocap data to acquire human motion priors without manual reward engineering. The paper also demonstrates downstream utilities, such as a real-time kinematic planner bridging motion tracking with interactive tasks like navigation. Multiple research teams have already started building upon SONIC.
Related event: NVIDIA Unveils SONIC for Humanoid Robot Control(3 posts)→
More from Embodied
- IHMC's HexRunner hits 30+ mph using spring-loaded legs instead of wheels — TinfoilTricorn · 2026-08-13
- Workers in India reportedly paid to film manual labor for robot training — Polymarket · 2026-08-13
- Black Forest Labs Teases FLUX 3: One Model for Image, Video, Audio, and Robotics — bennash · 2026-08-13
- Dev Integrates Vision for Local Agent Training, Explores Water Wave Computing — cephaloform · 2026-08-13
- AtlasVLA: Vision-Language-Action Model with Persistent Memory — CASIA-IVA-Lab · 2026-08-13
- Agility Robotics exec: Backflips easy, picking up a pen is harder — jonstephens85 · 2026-08-13