# Open-source projects train a humanoid's software in 3 minutes, in simulation, according to Ros Shinkle

> Training the software that controls a humanoid robot's whole body used to take a cluster of graphics cards. Open-source projects, built on one another from a research paper, now do it with a single card, according to Humanoids Daily.

Published on 2026-11-17

Ros Shinkle, founder and CEO of Tuesday Labs and a former roboticist at Boston Dynamics, sums up the change on X. According to her, training whole-body control software took a week on 128 graphics cards before 20 October. It now takes 3 minutes on a single card.

It all starts with SuperMimic, a paper published on arXiv on 20 October by a UC Berkeley team. Its authors say 3 hours on one graphics card gives better results than a 7-day baseline training run on 128 cards.

The paper came without code. The next day, Chong Zhang, of ETH Zürich, published an unofficial reimplementation. InstantTrack, BlinkMimic and XMimic were then built from it.

The times reported by BlinkMimic and XMimic are measured in simulation, on a benchmark of 40 motion sequences. Humanoids Daily points out that none of these projects reports a test on a real robot.

According to Humanoids Daily, training that takes a few minutes lets a small team test far more variants of its control software in a day, without renting a compute cluster.

## Sources

- [Humanoids Daily](https://www.humanoidsdaily.com/news/supermimic-blinkmimic-xmimic-humanoid-motion-tracking-minutes)

---

HTML version: https://vuntum.com/en/news/open-projects-cut-unitree-g1-humanoid-training-to-minutes-in-simulation

Factual data: CC BY-NC 4.0. Attribution: “Source: Vuntum (https://vuntum.com)” with a link to the page quoted. Editorial text: all rights reserved, short quotation with attribution.
