On the closing night of the World Humanoid Robot Games 2026, the organisers announced more than 2,500 hours of real-world robot data, free of charge to companies, universities and research institutions. A month later, on 27 September, we could not find anywhere to download a single hour of it. The announcement is real and so is the promise, but until the data appears, the humanoid robot dataset anyone can use today is a different one, and one of the closest to the Games comes from a company that won medals there.
What is in the World Humanoid Robot Games dataset?
The fullest English description comes from CGTN, the channel of China Media Group, one of the Games’ hosts, in its report on the closing ceremony of 26 August: “organizers released a dataset containing more than 2,500 hours of real-world operational data gathered from training and competition. The dataset covers 12 application scenarios, more than 100 skills and over 10,000 detailed tasks, and will be made available free of charge to companies, universities and research institutions.”
That description has two tenses. The dataset was “released” at the ceremony, and it “will be made available” to users. The report gives no download address, no licence, no file format and no size in bytes. It does not say who will host the files or how a university would apply for them. The same report carries the closing results we set out in our piece on the 2026 results.
The contents, as described, would be unusual. The two public datasets we cover below were each recorded by one company, on one robot or one motion capture system. Data “gathered from training and competition” at the Games would come from many teams and many makers, working through tasks the organisers designed and scored, such as the hotel, library and factory posts we describe in our piece on the scenario events. Each scenario run at the Games was also scored as autonomous or remote-controlled, a label that would make the data more useful if it survives into the files.
Can you download the World Humanoid Robot Games dataset?
Not yet, as far as we can find. On 27 September we checked the three places a release would most likely appear. The official site’s lists of news, announcements and documents held no item about a dataset, and its announcements list ends with notices from the final day. A search of Hugging Face, the main public host for machine learning datasets, for the Games’ name and its abbreviations returned nothing. A search of ModelScope, the Chinese equivalent, returned nothing relevant either.
That does not mean the dataset does not exist. It may be released through an application process, as some industrial datasets are, or on a site we did not check. It does mean that a researcher who heard the announcement cannot yet act on it, and that the size and contents above are the organisers’ description, not something anyone outside can verify.
When the files appear, a few details will decide their value. The licence sets whether a company can train a commercial robot on them. The format sets how much work it takes to load them. And each episode needs to record the robot model and the control mode, the Games’ most distinctive label.
Which humanoid robot datasets can you download now?
Two public datasets sit close to the Games in what they record. The first is AGIBOT WORLD 2026 on Hugging Face, published by AgiBot, the company that claimed the most gold medals in Beijing, as we describe in our piece on the medal table. Its card says the data is “collected from 100% real-world environments, covering commercial spaces, home, and other general-purpose scenarios”, on the AgiBot G2 robot, the same model AgiBot entered in the scenario events. It is licensed under CC BY-NC-SA 4.0, which allows research use but not commercial use, and follows the LeRobot v2.1 format. The repository was created on 11 March 2026, is not gated, and uses about 14 terabytes of storage according to the Hugging Face API. The card does not mention the Games. China News Service, the state news agency, names AgiBot among the companies behind the Games dataset, and between 17 and 26 September AgiBot added four new datasets to Hugging Face, each behind an access request. None of their names mentions the Games.
The second is MotionDecode, from the Shanghai motion capture company CHINGMU. Its card announces “1000 hours of robot-ready motion trajectories retargeted to the Unitree G1 humanoid”, recorded at 120 frames a second and supplied as CSV files. This is human movement mapped onto a robot, not the robot’s own experience of doing a job. The card calls the dataset “fully open and publicly accessible” and also has a request-access button for the complete set, and CHINGMU’s own launch page says the full dataset is available by application. The sources disagree on how open the complete set is; the sample files on Hugging Face download without a login. Its licence permits “academic research, personal study, prototyping (non-commercial)”.
| Dataset | On 27 September |
|---|---|
| Games dataset, from the organisers | Over 2,500 hours announced; licence not stated; not found online |
| AGIBOT WORLD 2026, from AgiBot | G2 robot data; CC BY-NC-SA 4.0; open download |
| MotionDecode, from CHINGMU | Over 1,000 hours of human motion; non-commercial terms; open on Hugging Face, full set by application according to CHINGMU |
Neither public dataset allows commercial training without further permission, which is the gap an open Games dataset could fill.
What a humanoid robot dataset from the Games would add
Robot learning runs on two kinds of data: simulation, which is cheap and endless, and real-world runs, which are slow and costly to collect. AgiBot’s card says its real-world scenes also have digital twins in its GenieSim simulation project. The simulation half is something anyone with a laptop can start on, as we found when we installed and ran MuJoCo on a Mac. The real-world half is what the Games’ 2,500 hours would supply at scale, from many makers at once.
It would also be data with results attached. The organisers publish a result sheet for every scenario final, with each team’s score and control mode. A dataset that lets a researcher match a run to its sheet would show not only what the robots did but whether it counted. Makers could then be compared on the same tasks, which is more than any single company’s dataset can offer and more than the shipment claims we read in our piece on how many humanoid robots shipped.
The organisers promised the data free to companies, universities and research institutions. The test of that promise is a download page, and we will add its address here when one appears.