Collect and manage data

Create a collection task

1. Log on to the Embodied Intelligence Platform.

2. In the left-side navigation pane, expand Tasks and click Data Collection.

3. Click Data Collection.

4. Configure the following parameters in the right-side panel:

   Basic information

   Parameter	Required	Description
   Task Name	Required	The name of the collection task.
   Dataset Name	Required	The name of the dataset where the collected data is stored, such as my_robot_task_001.
   Dataset Type	Required	Default: LeRobot.
   Device	Required	Select a registered online device from the drop-down list.
   Assign User	Required	Select the data collector who performs the collection.

   Task details

   Parameter	Required	Description
   Task Goal	Optional	The goal description of the collection task.
   Description	Optional	The detailed description of the task.
   Steps	Optional	The operation guide.
   XR Mode	Optional	Select an XR collection mode from the drop-down list.

   Other settings

   Parameter	Required	Description
   Status	-	Default: Collecting.
   Priority	-	Default: Medium.
   Tags	Optional	Add custom tags in Key-Value format for subsequent filtering.
   Repository ID	-	Default: lerobot/pusht.
   Estimated Duration	Optional	The estimated duration of data collection.
   Expected Collection Count	Optional	The expected number of episodes to collect.
   Dataset Storage Path	Optional	The custom storage path.

5. Click OK.

Execute data collection

After a collection task is created, the device executes the collection process through the adb-robot-client. The collection process consists of five steps:

1. Establish connection: The platform sends a connection command to the device through MQTT.

2. Confirm connection: The device confirms that the connection is established, and the status changes to READY.

3. Start recording: The platform sends a recording command, and the device starts collecting operation data (joint angles, torque, camera feeds, etc.).

4. Collection loop: Click stop recording to finish the current episode, and then click start recording to begin the next collection round.

5. Disconnect: The MQTT connection is disconnected after the collection is complete.

Upload data

After the collection is complete, you can upload data to the platform:

1. In the left-side navigation pane, click Data Upload.

2. View the upload progress and status of each task in the upload list.

The upload process supports filtering by episode review status and provides real-time progress monitoring.

Register a dataset

In addition to collected data, you can register existing datasets to the platform.

1. In the left-side navigation pane, click Custom Datasets.

2. On the Custom Datasets tab, click New Dataset.

3. Configure the following parameters:

   Basic information

   Parameter	Required	Description
   Dataset Name	Required	The name of the dataset.
   Data Format	Required	Select a format from the drop-down list: LeRobot 2.x, LeRobot 3.x, Unitree, EgoVerse Zarr, HDF5, or Other.
   Tags	Optional	Add tags in Key-Value format.

   Data source

   Switch between tabs to select a data source. The platform supports the following sources:

   Source	Description
   Local File	Drag or select a local file to upload.
   OSS	Import from Object Storage Service (OSS). You must enter the OSS Endpoint, Bucket, data path, AccessKey ID, and AccessKey Secret. For more information, see Upload data by using AnalyticDB lake storage OSS.
   Development Workspace	Import from the working directory of a running development machine. Select a development environment and specify the data path (default: /data/).
   Hugging Face	Coming soon.
   ModelScope	Coming soon.

4. Click OK.

   The Datasets page is organized into three tabs: Custom Datasets (manually registered datasets), Collection Datasets (datasets generated by collection tasks), and Annotation Datasets (datasets generated by annotation tasks).

Review data

Data review is used to evaluate the quality of collected data and supports per-episode review.

1. In the left-side navigation pane, click Data Review.

2. In the review task list, click the name of the target task to go to the details page.

   Dataset Viewer

   The details page provides a Dataset Viewer feature that displays dataset metadata (total number of episodes, data format, total frames, and total duration) and a frame-level data table, including all camera views, actions, and states.

   Episode review

   On the Episodes tab, each episode is displayed as a card with a thumbnail and duration. Click the action button to perform a review:

   Parameter	Description
   Review Result	Pass or Fail.
   Rating	1 to 5 star rating.
   Review Notes	Optional. Additional remarks.

Format conversion

When you need to convert a dataset from one format to another (for example, from EgoVerse Zarr format to LeRobot v2 format), you can use the data processing feature. Format conversion is based on the Ray distributed engine and supports efficient conversion of large-scale datasets.

1. In the left-side navigation pane, click Data Processing.

2. On the Format Conversion tab, click Create Format Conversion Task.

3. Configure the following parameters:

   Parameter	Required	Description
   Source Dataset	Required	Select the dataset to convert.
   Output Dataset	Required	The name of the converted dataset.
   Conversion Mode	-	Standard conversion (default). Custom conversion is coming soon.
   Format Type	Required	The target format, such as lerobot_v2. The following conversion directions are currently supported: Unitree → LeRobot v2 LeRobot v2 → LeRobot v3 EgoVerse Zarr → LeRobot v2

   Expand Advanced Configuration to configure the following parameters:

   Parameter	Description
   Output Mode	video (default) or image.
   Robot Type	One of four configurations for Unitree G1 (Inspire / Dex1 / Dex3 / Brainco).
   Execution Mode	Default: ray (Ray distributed execution).

4. Click OK.

After a conversion task is submitted, you can monitor the task execution progress through the Ray Dashboard.

Related documentation

• Manage devices and robot access

• Annotate data (Human-in-the-Loop)

• Upload data by using AnalyticDB lake storage OSS