Datasets

Datasets holds your recorded robot data. You upload a recording, BotManifold processes it into a training-ready dataset and catalogs each episode, and you browse the episodes and their signals in the web app. Datasets are private to your account.

Get data in

• From a robot: record with the teleop bridge and upload from the Studio Teleop panel, or run botmanifold bridge upload <session-dir>.
• From files: upload a recording in one of the supported formats: ROS 2 bag, MCAP, CSV or JSON, HDF5, or LeRobot v3.

Large files go straight to cloud storage over a resumable upload, then processing starts.

Processing

For each upload the service:

1. Parses the recording into episodes.
2. Optionally labels task phases with a vision-language model and validates the physics by replaying the actions in MuJoCo.
3. Computes quality metrics: completeness, temporal consistency, action diversity and reward coverage.
4. Exports the result (LeRobot v3, HDF5, RLDS or Open X) and summarizes every episode: steps, duration, task, quality score and action spread.

A dataset is pending after you start the upload, processing while it runs, then ready or failed. A failed dataset keeps the reason.

Browse and curate

Open Datasets in the web app to see your datasets and open one to see its episodes and quality report. An episode page shows the recorded signals:

• Time series of joint states, actions and rewards, with a shared timeline
• Trajectory: the robot model replaying the recorded joints in 3D
• Data table, which you can export as CSV

Under Pipeline → Curate you can exclude, include or relabel episodes before you train.

From the command line

botmanifold datasets list
botmanifold datasets quality <dataset_id>
botmanifold datasets episodes <dataset_id>

The SDK exposes the same through client.pipeline: list_datasets, get_dataset_quality, list_episodes and exclude_episodes. See the API reference.

Open datasets

The Explore page is a curated list of open-source robotics datasets (LeRobot format, with the task, robot, cameras, and episode counts). It does not need an account and is separate from your own uploads.