Quickstart

This path takes one recording from upload to an exported LeRobot v3 dataset. Everything runs on the same immutable dataset version, so what you export is exactly what you reviewed.

1. Sign up and create a project

1. Sign up at the console and create your team.
2. Open Projects and choose New project. Enter a project name and pick the recording kind: Robot (teleoperation, arms, grippers) or Human (first-person human demonstrations).
3. Choose Create. The project is where recordings, workflows, episodes and dataset versions live.

2. Upload a recording

In the console, open the project, go to Recordings and upload a folder or file. Large files go directly to object storage and can resume after an interruption.

To upload from a terminal, install the CLI and sign in. The CLI binaries are embodiflow and ef. Pass your API base URL, because the default points at a development API.

embodiflow auth login --base-url https://api.embodiflow.com
embodiflow whoami --json
embodiflow projects list --json

Then upload a folder into the project. Use the project id from the list above. When --source-type is omitted it is derived from the project recording kind.

embodiflow uploads create --project-id <project-id> --source-path ./recordings

You can also create a named recording with a team/project reference:

embodiflow recordings upload ./recordings/rec-0001 --project <team>/<project> --name rec-0001

If a transfer is interrupted, run the same command again with --resume.

3. Watch the workflow run

Submitting an upload starts the project workflow: it standardizes the recording, runs the quality checks and, where the project needs them, privacy and perception steps. Open Workflow, then Runs, to follow each step. A step that is not applicable or skipped is shown as such, not as passed.

embodiflow workflow runs list --project-id <project-id>
embodiflow workflow runs wait <run-id>

4. Review episodes

When the run finishes, open Episodes. Each episode is one task attempt cut from a recording. Check the quality results, open an episode in the replay, and exclude any episode that should not be trained on.

5. Create a dataset version

Open Versions and choose Create version. Select the episodes that pass checks, a split, labels and preprocessing. A version is immutable: later changes create a new version number instead of rewriting this one.

6. Export and download

Export the version as LeRobot v3 from the version page, or download it in one step with the CLI. The reference is <team>/<project>@<version number>.

embodiflow versions download <team>/<project>@1 --format lerobot-v3

The files are written to the local cache (~/.cache/embodiflow/datasets by default; use --dir to change it). To control each step yourself:

embodiflow exports create --project-id <project-id> --version-number 1
embodiflow exports wait --project-id <project-id> --version-number 1 --export-id <export-id>
embodiflow exports download --project-id <project-id> --version-number 1 --export-id <export-id> --target-dir ./export

Next

• CLI reference: every command group with an example.
• REST API and personal access tokens for automation.