Zhejiang University and Unitree Open-Source 5 Million Verified Motions for Robots
To teach a robot to squat or mop a floor, researchers usually record people wearing motion-capture suits. That is expensive, and it yields little data. A lab at Zhejiang University and Unitree instead built Movens-5M from ordinary internet videos: 5 million robot motions totaling more than 10,000 hours. Each motion was run in a simulator, and only the ones a robot can actually perform were kept. For developers, it is an open base for pretraining that saves them from recording data from scratch.
AI-processed from Jiqizhixin (机器之心); edited by Hamidun News
To teach a robot to squat, mop a floor, or box, researchers usually record the motions on real people in motion-capture suits and then test them on physical hardware. This is expensive and slow, so there is far less data for robots than there is text for language models. The joint embodied intelligence lab of Zhejiang University and Unitree has proposed a workaround: take ordinary internet videos of people doing things and turn them into robot motions.
The result is Movens-5M, an open dataset of 5 million motions totaling more than 10,000 hours, spanning 297 action types and 16 scene categories. If you build control models for humanoid or quadruped robots, this is a ready-made pretraining base that you do not have to record yourself.
What exactly was released
The name Movens-5M is simple to decode: five million motions. Each of them was executed in a simulator, which means it was checked for feasibility. In the Chinese headline of the announcement, the release was called a potential "GPT moment" for embodied intelligence, that is, for robots that act in the physical world. That is a bold claim, but the dataset itself deserves attention.
The motions are split into 297 types: walking, squatting, mopping floors, boxing, and others. The scenes fall into 16 categories: home life, fitness workouts, outdoor leisure. Simple arithmetic shows that the average sample lasts about 7 seconds. These are short, self-contained episodes rather than long continuous recordings. The project involved Zhejiang University, the lab together with Unitree, which is known for its quadruped and humanoid robots, and other teams.
How it works
The idea will be familiar to anyone who has watched videos on hosting platforms. A person mops a floor, does morning exercises, hits a punching bag. From such a video, you can extract the body's trajectory and transfer it to a robot. But a human body and a robot body differ: they have different proportions, joints, and limits on angles and balance. As a result, some of the "translated" motions are physically impossible for a robot, and it would simply fall over.
This is where the verification step matters. According to the authors, every motion was executed in a simulator. Infeasible samples were filtered out, and what remained are motions that the robot model can actually reproduce. This matters for training: the neural network does not waste effort memorizing motions that cannot be repeated.
What to do with it in practice
First, pretraining. You can teach a model basic skills on Movens-5M and then fine-tune it on a small amount of your own recordings for a specific task and a specific robot. This is cheaper than collecting everything from scratch.
Second, fair comparison. When a dataset is open, different groups can train their models on the same material and compare results. In robotics this is still rare: more often, each company sits on closed data.
Third, a signal for founders. Chinese labs are betting on scaling data, as with language models, and are publishing it openly. Anyone collecting closed data should take into account that the open alternative has become noticeably larger.
Caveats
The available part of the announcement lacks information that is needed to assess the dataset. It is unknown which robots and which simulators were used to verify the motions, what license the data carries, and what results models trained on it have shown. Simulation remains an approximation of physics: a motion that is feasible in a simulator may behave differently on a real robot because of friction, backlash, and latency. In addition, 297 action types and 16 scene categories cover many everyday situations, but that is not yet all of real life. This is especially noticeable in tasks where a robot has to pick up and rearrange objects, not just move its own body.
So it is too early to call Movens-5M a "GPT moment." First, we need to see how much better models pretrained on this data perform on tasks with real robots. But an open 10,000-hour dataset with feasibility checks is already a useful resource. If you work with humanoid or quadruped robots, it is worth downloading the dataset, studying its license and format, and testing whether it cuts your data-collection costs.
- Nov 20, 2026Four Hours on an RTX 5090 Cost 348 ₽, or From 1 730 ₽ With a One-Day Minimum
- Nov 20, 2026In China, Job Postings for AI-Exposed Professions Fell 10 Percentage Points Faster Than the Rest
- Nov 20, 2026Suno adds Visuals: turn a song into a clip of up to 15 seconds with lip-sync, for separate credits
- Nov 20, 2026HeyGen claims 84% of viewers mistook its Avatar 6 AI avatar for a real person
Key Questions Answered
What exactly was released?
The name Movens-5M is simple to decode: five million motions. Each of them was executed in a simulator, which means it was checked for feasibility. In the Chinese headline of the announcement, the release was called a potential "GPT moment" for embodied intelligence, that is, for robots that act in the physical world. That is a bold claim, but the dataset itself deserves attention. The motions are split into 297 types: walking, squatting, mopping floors, boxing, and others.
How it works?
The idea will be familiar to anyone who has watched videos on hosting platforms. A person mops a floor, does morning exercises, hits a punching bag. From such a video, you can extract the body's trajectory and transfer it to a robot. But a human body and a robot body differ: they have different proportions, joints, and limits on angles and balance. As a result, some of the "translated" motions are physically impossible for a robot, and it would simply fall over. This is where the verification step matters.
What to do with it in practice?
First, pretraining. You can teach a model basic skills on Movens-5M and then fine-tune it on a small amount of your own recordings for a specific task and a specific robot. This is cheaper than collecting everything from scratch. Second, fair comparison. When a dataset is open, different groups can train their models on the same material and compare results. In robotics this is still rare: more often, each company sits on closed data. Third, a signal for founders.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.