# NablaMimic preprint reports faster humanoid motion training on one GPU

_Published Monday, November 2, 2026 at 12:20 AM EST · Robotics, AI, Science · Latest · Tier 2 — Notable_

Researchers report that NablaMimic trains a humanoid motion-tracking policy to a score of 0.82 in 20 minutes, compared with 1.2 h for a leading world-model method. Their preprint tests one policy across all 40 LAFAN1 motion clips on one GPU and reports a score of 0.874 after 1.8 h.

The method limits unstable gradients and reuses simulated sequences to train actions. Its NablaEngine simulator computes contact dynamics and their gradients for the Unitree G1. The researchers release both tools as open-source infrastructure; the reported results concern simulated motion tracking.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.41654)

---
Canonical: https://techandbusiness.org/newswire/q68QyWxIOuLVI0F7xjuVR3
Published: 2026-11-02T05:20:44.593Z
Story chronology: 2026-11-02T05:00:00.000Z
Retrieved: 2026-11-02T07:51:26.119Z
Publisher: Tech & Business (techandbusiness.org)