NVIDIA GEAR's SONIC whole-body control ported to humanoid Asimov, running at 2.6ms per tick on ARM
freelerobot · x · 2026-10-04
Menlo Research, with a port by UFBots, deployed NVIDIA GEAR's open-source whole-body control policy SONIC on the humanoid robot Asimov. SONIC encodes reference motion into quantized tokens and decodes them with robot state into joint targets; public checkpoints are 22.5M–50M parameters.
Key results:
- SONIC for Asimov (13.8M parameters, one encoder + decoder) runs a full inference tick in 2.6ms on two ARM A76 cores of Asimov's RK3588 (4.2ms single-core), faster than the typical 50Hz control tick
- Full GEAR-SONIC also fits on the same CPU: 4.5ms for the original release, 9.4ms for the 49.7M-parameter v1.1
- SONIC gets a dedicated core, leaving the 200Hz motor loop untouched
The work shows lightweight open-source whole-body control can run in real time on cheap on-device compute.
More from Embodied
- 45 years of Marc Raibert: from one-legged hoppers to SoftBank buying his new lab twice — lukas_m_ziegler · 2026-11-20
- Meta to price Muse Charm AI device at $300 with up to a year of free Verizon 5G — Scobleizer · 2026-11-20
- Meta's Muse Charm AI gadget reportedly priced at $300, shipping by December — The Verge AI · 2026-11-20
- Nest doorbell catches a clawed robot autonomously tidying a yard — Buderus69 · 2026-11-20
- NVIDIA showcases developers building simulation apps with frontier AI agents like GPT-6 Astra — liu_mingyu · 2026-11-20
- NVIDIA and Microsoft bring Vera Rubin NVL72 to Azure and Nemotron 3.5 to Foundry — NVIDIA Blog · 2026-11-20