An algorithm is the code that one workflow step runs. The Algorithms page lists the built-in algorithms the platform provides, and the team algorithms your team has registered.
Who can do this: Team roles Owner, Admin, Researcher and Viewer can view the list. Creating a team algorithm needs the algorithm write permission: Owner, Admin, Researcher.
| Area | What it does |
|---|---|
| ① Built-in and Custom | Switch between the platform algorithms and your team's algorithms. |
| ② New algorithm | Create a team algorithm. See Team algorithms. |
| ③ More actions | Open the menu of more actions for this page. |
| ④ Algorithm list | The built-in algorithms, grouped by stage. Choose a name to see its details. |
Built-in algorithms
Built-in algorithms are read-only. You cannot edit, publish, check or run them from this list. They run as steps in a workflow.
| Stage | Algorithm | What it does |
|---|---|---|
| Ingest | Ingest | Reads the upload and records its data kind, format, duration, cameras and available pose signals. |
| Pre-check | Extract metrics | Measures a raw recording (ROS bag, MCAP or video) and writes quality metrics before standardizing. |
| Pre-check | Quality gate | Checks quality metrics before standardizing. A failed check stops the recording. |
| Standardize | Standardize | Converts a robot or human recording to a LeRobot v3 recording at 30 Hz, using its rig profile. |
| Standardize | Standardize video | Converts a first-person video to a LeRobot v3 recording at 30 Hz. Audio is dropped. |
| Privacy | Anonymize | Blurs faces and license plates in every camera of a recording. Needs a GPU. |
| Perceive | Camera pose (VIO) | Estimates head pose for every frame from the head stereo pair and head IMU. |
| Perceive | Hand pose | Finds 3D hand keypoints from both head cameras. |
| Perceive | Sensor calibration | Measures stereo rotation and the camera to IMU time offset from each recording. |
| Perceive | Stereo depth | Computes metric depth for the left head camera from the stereo pair. Needs a GPU with at least 9 GB of memory. |
| Annotate | Propose episodes | Proposes episodes with start, end, instruction and skill using a vision-language model. Proposals are not reviewed. |
| Compose | Compose recording | Merges the anonymized recording with hand, head pose and depth results into one recording. |
| Save results | Persist result | Saves the quality report and the episodes of one recording. |
| Check | Evaluate quality | Turns step metrics into the recording's quality report. |
| Check | Publish report | Publishes the quality report for the batch. |
| Check | Validate standardized data | Checks that the standardized recording meets the minimum format. |
| Check | Visual metrics | Measures exposure, blur, frozen frames and camera dropout on sampled frames. |
| Export | LeRobot export | Exports a dataset version as a LeRobot v3 archive. |