Skip to content

Dataset Preparation - Recording

Before Recording

1. Open the Web UI

a. Access the Web UI in Your Browser

Please enter the IP address of the PC running Physical AI Server (physical_ai_server_bringup.launch.py or ai_server) into your web browser.

TIP

If you are running the Physical AI Server and opening the web browser on the same computer,
you can simply enter http://localhost

For example:
  • If the PC running Physical AI Server has IP address 192.168.0.105
    Access http://192.168.0.105 as shown below:
Web UI

Enter the IP Address

  • Once accessed, the following screen will appear with the Disconnected status.

    Web UI

    Initial Screen

2. Select the Robot Type

On the Home page, select the type of robot you are using.

Web UI

After clicking the Set Robot Type button, the status will change to Connected.

Web UI

Connected Screen

3. Go to Record page

INFO

You cannot access Record page unless a robot type has been selected on the Home page. Please ensure that the robot type is selected before proceeding.

After the status changes to Connected, you can access the Record page using the button on the left sidebar, as shown below.

Web UI

Record Screen

The Record page is divided into three main sections:

  • Image Streaming Area (Center): Displays real-time camera streams from the robot.

  • Task Info Panel (Right side): Enter task name, task instruction, time values, etc.

  • Control Panel (Bottom): Start/stop recording and manage data collection.

The selected robot type is also displayed in the top left corner.

4. Visualize RGB Images

The image stream is displayed automatically upon entering the Record page. You can remove the current stream and select a different one as needed.

To change the image topic:

  1. Click the + button in the Image Streaming Area.
  2. Choose a topic from the popup window.
Web UI

5. Enter Task Information

Fill out the task-related fields in the Task Info Panel, as shown below. For more information about what each field means in the Task Info Panel, please refer to 👉 Task Information Field Descriptions below.

Web UI

Filled Panel

👉 Task Information Field Descriptions
ItemDescription
Task NameIt is used for the folder name of the dataset
Task InstructionIt is a sentence that instructs the robot what action to perform, such as "pick and place object".
In multi-task mode, you can add Task Instructions by clicking the Add Instruction button.
Push to hubIf you want to push the dataset to the Hugging Face Hub, check this box. This allows sharing and training.
To push to the Hugging Face Hub, you need to:
1. Have a Hugging Face account
2. Have the necessary permissions to push to the repository
Private ModeOnly available when "Push to hub" is checked. Keeps your dataset private on Hugging Face.
User IDYour Hugging Face account username, also used as the folder name for the dataset.
- If not using Hugging Face, you can use any name.
- If using Hugging Face, click Load to select from registered User IDs.
- If no User IDs or want a different one, click Change and enter your Hugging Face token.
- The Hugging Face token must have both read and write permissions enabled.Web UI
- If you have a registered account but haven't checked "Push to hub", you can either:
1. Load a registered User ID
2. Manually enter any User ID
FPSFrame rate for dataset recording. Recommended value is 15.
TagsKeywords to categorize and organize your dataset in the hub. Multiple tags can be added. Useful for searching or filtering later.
Warmup TimeDuration (in seconds) to warm up the robot before starting the recording
Episode TimeDuration (in seconds) to record each episode
Reset TimeTime allocated (in seconds) for resetting between episodes
Num EpisodesTotal number of episodes to record
Optimized SaveIf enabled, uses RAM for faster dataset encoding processing

You are now ready to start recording your dataset.

Start Recording

To begin recording, use the Control Panel located at the bottom of the Record page:

Web UI

Step 1: Click the Start Button

Click the Start button to begin the recording session. The system will automatically:

  • Warm up the robot for the specified Warm-up Time
  • Record each episode for the specified Episode Time
  • Wait for the specified Reset Time between episodes
  • Repeat the above steps for the specified Number of Episodes

Step 2: Monitor and Control During Recording

While recording is in progress, the following controls are available:

  • The Stop button saves the current episode in progress and stops the recording. If you press the Start button again, recording will resume from the next episode
  • The Retry button cancels the current episode and restarts recording for that episode
  • The Next button ends the current episode early and moves to the next episode
  • The Finish button ends the recording session and saves the current dataset, regardless of the remaining number of episodes

INFO

  • The current recording stage is displayed in the control panel:

    • 📍 Ready to start — Standby mode before recording begins
    • 🔥 Warm-up in progress — Robot is warming up
    • 🔴 Recording in progress — Capturing data
    • 🏠 Reset in progress — Reset time between episodes
    • 💾 Saving... — Encoding and saving the episode

TIP

  • 🔊 A beep alarm will sound at the start of each recording to indicate the timing.

After Recording

  • The dataset will be saved on the host machine.
  • If "Push to Hub" is enabled, the dataset will be uploaded to Hugging Face.
  • You can access the recorded dataset at the following path from from the host environment or Docker container:

USER PC

bash
<your_workspace>/physical_ai_tools/docker/huggingface/lerobot/

or

USER PC 🐋 PHYSICAL AI TOOLS

bash
${HOME}/.cache/huggingface/lerobot/

AI Worker and AI Manipulator released under the Apache-2.0 license.