Kinetic Blocks has opened a public catalogue of open robotics datasets. It is live today at kineticblocks.com/open-data, and it needs no login.
The catalogue indexes 27 publicly released datasets of robot video and first-person human video, close to 19,000 hours of footage as their publishers state it, from DROID and Open X-Embodiment through to EPIC-KITCHENS, AgiBot World and Egocentric-10K.
Anyone who has put together a training set from open sources will know the problem. The datasets themselves are not hard to find, but they are scattered across university servers, cloud buckets and repository pages, each describing itself in its own vocabulary and each attaching its own licence, frequently as a code rather than an explanation. Comparing two of them on hours, task coverage or commercial terms means opening a dozen tabs and reading carefully enough to be confident you have understood what you are permitted to do. Every team that goes looking ends up doing that work again.
Each dataset in the catalogue therefore has a card in the same format as a listing on the marketplace, with the hours first, then the episode count, the task, the environment, the modalities, where the footage was captured and where the files are hosted. Every card states the licence in plain words rather than a code: commercial use permitted, training permitted but no resale, or non-commercial only. A search box and filters narrow the set by footage type, licence, access, task and environment.
"We see a need to collect all open data sets in one place. Both to support the community but also to help comparing the sets. With time, Kinetic Blocks aims to do quality control and usability checks on all datasets on our platform," says Lars-Fredrik Forberg, CEO of Kinetic Blocks.
Where the files actually live
Kinetic Blocks hosts none of these files. Each dataset page restates the publisher's licence, links to the publisher's own page and terms, and hands you to the host, whether that is Hugging Face, a Google Cloud bucket or a university server, with the download command ready to copy where one exists.
The figures on the cards are the publishers' own. Where a dataset reports frame counts rather than hours, the hours are derived and marked as such. None of these datasets has been verified or graded by Kinetic Blocks, and every card says so.
That distinction is why we published the catalogue in the first place. On the marketplace, a commercial listing is checked against its own files and carries a KB Quality grade before it goes on sale. The open catalogue sits alongside it, so that a team choosing training data can weigh the free options against the verified ones in one place and on comparable terms. Which of the two makes sense will depend on the task, the budget, and how much of the quality checking a team wants to do itself.
What comes next
We are checking licences on the next batch of datasets now, and over time we want to run open data through the same verification and grading as the marketplace listings.
Browse the catalogue at kineticblocks.com/open-data. If you publish a dataset and would like it indexed, or if a card needs correcting, write to sales@kineticblocks.com.