# BlinkMimic trains simulated humanoid motion policy in 3.4 minutes

_Published Friday, November 13, 2026 at 1:02 AM EST · Robotics, Science · Latest · Tier 2 — Notable_

Researchers report training a simulated Unitree G1 humanoid to follow all 40 LAFAN1 motion clips in 3.4 minutes on a single RTX 4090. Their arXiv preprint presents BlinkMimic, a training recipe for a task they say has usually required hours of GPU time.

The recipe builds on earlier methods and adds a decreasing step size, resets guided by failures under deployment conditions, and GPU graph capture of the whole training iteration. The reported time is the mean of three seeds, measured until three consecutive checkpoints reached a normalized return of at least 0.82 with no failed clip. The demonstration is in simulation.

## Sources

- [arXiv Query: search_query=cat:cs.RO&id_list=&start=0&max_results=30](https://arxiv.org/abs/2611.16422v1)

---
Canonical: https://techandbusiness.org/newswire/CVZmJ4JiYIOwl7SCSrsEfi
Published: 2026-11-13T06:02:56.561Z
Story chronology: 2026-11-12T07:31:42.000Z
Retrieved: 2026-11-13T08:33:44.009Z
Publisher: Tech & Business (techandbusiness.org)