Research · 8 Oct 2026 · ⚠ Unconfirmed

VioLA preprint trains a humanoid policy mostly on human motion data

Purpose: sending the RankedRobot digest. Nothing else; never sold or shared for marketing.
What we keep: your email address, the date you signed up and the page you signed up on.
How long: until you unsubscribe, when the address is deleted. Before the first issue we email you once to confirm; addresses not confirmed within 30 days are deleted.
Who processes it: stored with Cloudflare, our host, which may process it outside your country. We'll name our email-sending provider in the privacy notice before the first issue.
Your choice: you don't have to subscribe; the only effect of saying no is that you won't get the digest. Unsubscribe any time.

The authors report that a policy predicting latent motion codes followed locomotion instructions on a real robot without task-specific fine-tuning.

A team led by Mert Albaba, with nine authors in all, posted VioLA to arXiv on 8 October. It is a preprint and has not been peer reviewed.

Rather than output raw joint commands, the policy predicts latent representations of body and hand motion, which pretrained controllers turn into movement. Motion encoders map human and robot motion into the same latent spaces, so human recordings can serve as training data.

The authors say the training pool holds about 140.6 million frames, roughly 93% of them human. On a real robot they report 100% success on locomotion instructions without fine-tuning, against 16.7% for GR00T N1.7 and 0% for Ψ₀, and 88.6% success on manipulation tasks. They say code and checkpoints will be released.

Sources

arXiv preprint (Albaba et al., arXiv:2610.12435)

Written by the RankedRobot news desk (AI-assisted, human-reviewed) from the sources above; checked 2026-10-11. Claims are attributed to whoever makes them. Spotted an error? corrections@rankedrobot.com