About
Robotics datasets don't share a layout. One is LeRobot v2, the next is LeRobot v3 with a dozen sub-datasets inside, another is RLDS shards, an MCAP recording, an HDF5 file or a Zarr store. Looking inside one usually means downloading it and writing a viewer for that one layout.
RoboPeek reads the layout for you. Paste a Hugging Face link to a public dataset, or any folder or file inside one. It works out what it is and opens the whole dataset in your browser.
What it opens
LeRobot v2 and v3, RLDS (TFRecord), MCAP (ROS 1, ROS 2, protobuf, JSON), HDF5, Zarr, and folders of camera videos
Repos that hold several datasets: each sub-dataset is listed and opens on its own
Datasets the Hub's own viewer can't preview, like the HDF5, Zarr and MCAP samples on the site
What you get for a dataset
Every episode in one list, with its length, task instruction and filters
All cameras in sync, each one switchable
Every recorded signal (actions, joint state, forces) charted frame by frame, with the task instruction on the same timeline
Stats on episode lengths, and Insights that find length outliers across all episodes
Flag episodes from the player or the lists, then export the flagged set as a LeRobot command
A link for every dataset and episode, so a colleague opens the same view
How it reads
Files stream straight from Hugging Face, and only the parts you look at are read
Nothing to install. Local TFRecord and MCAP files are read in the browser, not uploaded
Built to embed
The RLDS and MCAP readers and a React viewer are on npm under MIT. Put the viewer in a dataset catalogue, a labelling tool or a research dashboard, or use a reader on its own. datasets.bot is the first app built on it.
Limits today: public datasets only, and a few layouts still don't open fully (for example, camera frames stored inside Parquet files show joint data only). If yours doesn't open, open an issue on GitHub.
Built and maintained by VideoDB.