---
title: "Five million robot motions, each about seven seconds long"
date: 2026-11-20
category: Humanoids & Robotics
site: NeuroAI
canonical: https://neuroai.site/a/na-robot-movens-5m-1120
language: en
---

# Five million robot motions, each about seven seconds long

> Zhejiang University and Unitree open-sourced Movens-5M: five million simulated-verified robot motions averaging 7.2 seconds, plus scaling results from 1% to 100% of the data.

On November 18, the embodied intelligence lab run jointly by Zhejiang University and Unitree Robotics open-sourced a dataset called Movens-5M.

Five million robot motions, each verified by simulation execution. Another 15 million human motions accompany them. Total duration exceeds 10,000 hours.

The paper went up on arXiv on October 30 under the identifier 2610.41877. November 18 was the day the dataset and models were released together.

## Roughly seven seconds per motion

Ten thousand hours divided by five million clips gives 7.2 seconds each.

That matches the sample in the paper: a 7.25-second clip of someone wiping a door is cut into three phases — bending to wipe the lower panel from 0 to 4 seconds, straightening up from 4 to 6, reaching upward from 6 to 7.25. Each phase is annotated with which hand moves and where.

The dataset covers 297 fine-grained action types across 16 scene categories, from mopping and squatting to boxing.

## Five stages to turn a person into a robot

Raw video cannot be fed to a humanoid directly. People and the Unitree G1 differ in height, joint limits and balance strategy.

The team built five stages: split long videos along shot boundaries; filter clips for visual quality and motion content; annotate scenes and body-part motion over time; recover 3D human pose from frames; then fix jitter and retarget the motion to the G1 skeleton, checking in simulation whether the robot can execute it stably.

Only after that pipeline does the corpus contain 15 million human motions and 5 million robot motions. If any filter is too loose, the trained motion comes out shaky.

## Scaling, measured rather than asserted

With the motion expert fixed at 910.5 million parameters, the team raised training data from 1% to 100%, lifting deduplicated motion duration from 122 hours to 12,231 hours.

Text-to-motion retrieval rose from 12.72% to 62.57%. Video-to-motion error, measured as MPJPE, fell from 132.82 mm to 40.90 mm, a 69.2% reduction. Speech-motion matching climbed from 13.98% to 74.30%, and music beat alignment from 0.23 to 0.61.

A second experiment fixed the data and scaled only the motion expert, from 64.2 million to 910.5 million parameters. Validation loss dropped from 0.487 to 0.159, down 67.4%.

One finer result: raising data from 10% to 100% improved text-motion retrieval by 62.5% for the smallest expert and 151.1% for the largest.

When data is plentiful, bigger models absorb more of it. That has direct implications for anyone planning compute.

## Dancing is not the same as working

In the demos, typing "Dance some hip-hop" makes the G1 dance. Saying "Salute with your left hand" produces a salute. Music drives stepping and upper-body motion. A video demonstration is copied.

They are impressive, and they are short, single-task, controlled-environment clips.

The dataset has limits of its own. Every motion is retargeted to the G1, so a different body means redoing the pipeline. The gap between internet video and real hardware is measured with MPJPE, not eliminated. And because the source is public video, copyright and likeness questions are left open in the paper.

Five million is a hard number. It proves motion data can be produced at scale. It does not prove that a robot will run an eight-hour shift on a factory floor.

Between those two claims sit hardware reliability and on-site commissioning.

## Why size is the argument

The paper places its corpus against earlier efforts, and the gap is wide: the largest prior humanoid motion dataset it cites sits under 800,000 samples, against more than 5 million here, a factor of roughly 6.3.

That comparison is the point of the release. Robot foundation models have been data-starved in a way language models never were. Text on the internet is free and abundant; motion captured on real hardware is neither.

Turning public video into usable motion is one answer, and it is cheap at the margin. The expensive part is the pipeline, which has to run once and then scale. But video-derived motion carries the habits of the video: occlusions, camera shake, and human proportions that a 1.3-metre robot cannot reproduce.

The scaling curves suggest the approach works well enough to keep funding. What they do not show is whether the resulting policy transfers to a different robot, a different floor, or a task that takes twenty minutes instead of seven seconds.

*Disclaimer: This article draws on the paper (arXiv 2610.41877), the project page and reporting by Jiqizhixin. It is for reference only and does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-movens-5m-1120
Free to quote with attribution and a link to the original.
