NVIDIA's Metatron lets coding agents drive humanoid whole-body motion via a goal interface
carlosdponx · x · 2026-10-29
NVIDIA researchers unveiled Metatron, a promptable whole-body execution layer (a System 1 controller) for humanoids: a goal-based interface connects System 2 planners—including coding agents—to whole-body movement.
Key ideas:
- Agents specify "what and where": hand/foot/torso targets, travel velocity and heading, arrival times—constraining only what matters while the policy generates the rest
- One frozen G1 policy handles six behaviors: steering, teleoperation, sit & stand, recovery, physical guidance (free arms), and dense whole-body tracking
- Coding agents can write and refine goals to build new skills without retraining the policy
The team showed capable simulated climbing and sliding maneuvers. Project page, paper, and code are public; collaborators include Chen Tessler, Yifeng Jiang, and Xue Bin Peng.
More from coding & agent
- ChatGPT's dot can now text you via iMessage, with per-channel routing tips — nickbaumann_ · 2026-10-29
- Prevention beats repair: proactive AI saves ~$225K/year vs $120K for fix-it agents — louisdorard · 2026-10-29
- Product engineering in 2026: let Claude Code read the meeting transcript and build — curious_vii · 2026-10-29
- Anthropic researcher spends 3 straight workdays in Antigravity, vouches for Carbon — vedantmisra · 2026-10-29
- Microsoft launches Copilot harness in public preview within Foundry Agent Service — lee_stott · 2026-10-29
- Dev vibecodes a ComfyUI node that auto-archives and organizes all your generations — LaurentLaSalle · 2026-10-29