EmbodiFlowDocs
Get startedDataWorkflows and algorithmsTraining and deploymentDeveloper reference
  • Workflows and algorithms

Workflows

  • How workflows work
  • Build your own workflow
  • Runs and logs

Algorithms

  • Algorithm catalog
  • Write a custom algorithm

Algorithm developer guide

  • Your first algorithm
    • Standardize
    • Check
    • Transform
    • Enrich
    • Segment
    • Curate
    • Export
  • Image algorithms
  • Use your algorithm in a workflow
  • Interface reference
  • SDK reference
  • Command line
  • MCP tools
  • Troubleshooting

© 2026 EmbodiFlow

llms.txtOpenAPI JSON
  1. Workflows and algorithms
  2. Algorithms
  3. Algorithm catalog

Algorithm catalog

The algorithms your workflow steps run: the built-in ones, grouped by what they do, and the ones your team writes.

Every step in a workflow runs an algorithm. The Algorithms page is the catalog of all of them, so you can see what each one does, what it needs and what it produces before you put it in a workflow.

Built-in algorithms, by category

The Built-in tab groups the platform's algorithms by what they do, with a count for each group. Each row has a one-line summary and two chips on the right, such as "Input: Recording" and "Output: Features", so you can tell at a glance where it fits in a workflow.

CategoryWhat it doesBuilt-in algorithms
StandardizeTurns uploaded files into a LeRobot v3 recording.Canonicalize MCAP, Canonicalize video
CheckMeasures quality and writes checks.Raw-file metrics, Visual metrics
TransformChanges the recording itself.Anonymize
EnrichAdds new values to a recording, such as poses or depth.Sensor calibration, Hand pose, Hand pose (MANO), Head pose (visual-inertial), Stereo depth, Body pose (SMPL-H), Ego 3D visualization, Video preview, Episode success
SegmentFinds episodes in a recording.Propose episodes, Recording episodes
CurateDecides which episodes go into a new dataset version.Filter episodes, Split episodes
ExportPacks a dataset version into files to download.LeRobot v3 export

Two algorithms use a vision-language model: Propose episodes and Episode success. What they produce is a suggestion until a person reviews it, and the console labels it that way.

Typed inputs and outputs

Each algorithm declares the type of data it takes and gives: Raw files, Recording, Features, Checks, Episodes, Dataset version, Selection or Files. It also runs at one level: once per recording, once per episode, or once per dataset version. The workflow editor uses these types to connect steps for you and to refuse a connection that can't work.

The detail page

Click an algorithm to open its page. Inputs lists what it reads and which inputs are required, Outputs what it writes, and Parameters every setting with its default. These are the settings you change in the workflow editor. Algorithms that write checks or features list them by name too.

Some algorithms need a GPU. The summary in the catalog says so, and Resources on the right of the detail page gives the exact need, such as "1 GPU, at least 9 GiB" for Stereo depth. A GPU step waits in the queue until a GPU with enough memory is free.

Versions

Algorithms are versioned, and a workflow step uses one specific version, which its settings show. A built-in algorithm's page shows its Latest version. A team algorithm lists every saved version. A saved version never changes, so a published workflow keeps doing exactly what it did when you tested it.

Your team's algorithms

The Custom tab lists the algorithms your team has written, with their latest version. Search by name or code, and click one to see its page. In the workflow editor's library they sit in their category next to the built-in ones, marked Team.

To write one, click New algorithm. Team Owners, Admins and Researchers can create and save team algorithms. Write a custom algorithm shows how.

Next

  • Write a custom algorithm
  • Build your own workflow
  • Your first algorithm
Previous

Runs and logs

Next

Write a custom algorithm