EmbodiFlowDocs
Open console
Get startedDataWorkflows and algorithmsTraining and deploymentDeveloper reference
  • EmbodiFlow documentation

Get started

  • Quickstart
  • Key concepts
  • Console tour

Team

  • Team and members
  • Roles and permissions

Help

  • FAQ

© 2026 EmbodiFlow

llms.txtOpenAPI JSON
  1. Get started
  2. Get started
  3. Quickstart

Quickstart

Take one project from the first upload to a LeRobot v3 dataset you can download.

This walk-through takes you from an empty project to a dataset version you can download. Have one recording ready, for example a robot MCAP file or a first-person human recording. Upload recordings lists every format a project accepts.

Creating a project takes a team Owner or Admin. If someone has already added you to a project as an Editor, start at step 2.

1Create a project

In the sidebar, choose Projects, then New project at the top right. Pick the Recording kind: Robot for teleoperation and arms, Human for first-person demonstrations. The kind decides how uploads are processed and checked, so pick what the camera actually recorded. Name the project and choose Create.

2Upload a recording

Open the project and choose Upload data at the top right. Fill in Task, one short sentence about what the robot or person does, unless the project already has a default task. Then drag files or folders onto the page, or choose Add files. Each file or folder becomes one recording, and the right-hand column shows what will be created. Choose Start upload.

Keep the tab open until the upload finishes. Leaving the page stops it.

3Watch the run

You don't need to set anything up. As soon as the files arrive, the project's workflow starts on its own, and a green notice appears with View runs. The run page lists every step on the left, such as standardize, anonymize, quality checks and episode cutting, with the workflow graph beside it. When all steps finish, the run reads Succeeded.

4Look at the episodes and their quality

Choose Episodes in the project tree. A robot recording becomes exactly one episode. For a human recording, an AI model proposes where each task starts and ends and writes a task description; these stay marked AI-proposed, not reviewed until a person checks them. Choose Preview on a card to watch it.

Each episode also carries its quality result. Only Pass counts as a pass and shows green. Warning, Fail and Not checked are not passes, so open them in Quality when you want to know why.

5Create a dataset version

Choose Versions in the project tree, then Create version. The defaults include the episodes that are ready; you can also set the split into train, validation and test. Name the version and create it. From that moment it is frozen: the episodes and settings in it never change.

6Export and download

Open the version and choose Export. The format is LeRobot v3, with camera videos included; choose Start. When the export finishes, it appears under Exports with a load check, and Download gets you the pack. The same page shows how to load the version with Hugging Face tools.

Recordings that need privacy processing are exported only after anonymization has finished for them. The standard workflows take care of that in the anonymize step.

Next

  • Train a model on your version: choose Train model on the version page, or read Training.
  • Change what happens after an upload: Workflows.
  • Get your bearings in the console: Console tour.
Previous

EmbodiFlow documentation

Next

Key concepts