⚙ LeRobot Play Data Collection

LeRobot Play Data Collection#

Environment Configuration#

Environment Requirements Overview

Operating-system requirements

Architecture Ubuntu 22.04 LTS Ubuntu 24.04 LTS

x86_64 ✅ Supported ✅ Supported

ARM64 ✅ Supported ❌ Not supported
```
# Quick verification commands
echo "架构: $(uname -m)"
echo "版本: $(lsb_release -rs)"
```
Minimum hardware requirements
- CPU: Intel i5 or a processor with equivalent or better performance
- Memory: 8 GB RAM

Software Installation#

System Configuration and Python Version Matrix#

Architecture	Ubuntu Version	Python Version
x86_64	24.04	3.12
x86_64	22.04	3.10
ARM64	22.04	3.10

Installation Steps#

Determine your Python version from the table above, then run the corresponding commands:

# 1. Create a Python virtual environment
# Select the command for your system (choose one of three):

# Configuration 1: x86_64 + Ubuntu 24.04
conda create -y -n lerobot python=3.12

# Configuration 2: x86_64 + Ubuntu 22.04
conda create -y -n lerobot python=3.10

# Configuration 3: ARM64 + Ubuntu 22.04
conda create -y -n lerobot python=3.10

# 2. Activate the environment and install dependencies (all configurations use the following steps)
conda activate lerobot
conda install -c conda-forge libstdcxx-ng
conda install ffmpeg=7.1.1 -c conda-forge

# Install the AirBot core packages; replace <version> with the actual version
pip install airbot_hardware_py-<version>.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install lerobot_play-<version>.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install airbot_state_machine-<version>.whl -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install mmk2_kdl_py-<version>.whl -i https://pypi.tuna.tsinghua.edu.cn/simple

Quick Solutions to Common Issues

Issue 1: The wheel package platform is not supported

Error: ERROR: airbot_hardware_py-<*>.whl is not supported

Cause: The Python version does not match the system.

Solution: Recreate the environment with the corresponding Python version.

# Ubuntu 24.04 → Python 3.12
conda create -y -n lerobot python=3.12

# Ubuntu 22.04 → Python 3.10  
conda create -y -n lerobot python=3.10

Issue 2: The evdev package fails to install

Error: evdev compilation fails while installing lerobot_play-<*>.whl

Cause: GCC build tools are missing.

Solution:

# Install build tools and evdev
conda install -c conda-forge gcc_linux-64 gxx_linux-64 evdev
# Reinstall lerobot_play
pip install lerobot_play-<*>.whl

Hardware Configuration#

Hardware List#

Solution	Robotic Arm Configuration
Single-arm Play	1×AIRBOT Play + 1×AIRBOT Replay or 1×AIRBOT Play + 1×AIRBOT Play with E2
PTK	2×AIRBOT Play + 2×AIRBOT Replay
TOK2	2×AIRBOT Play + 2×AIRBOT Replay
TOK4	2×AIRBOT Play + 2×AIRBOT Play with E2
Quest3	1×Quest3 + 2×AIRBOT Play with G2
Note	All solutions can be freely combined with USB or RealSense cameras. The Intel RealSense D435i is recommended.

Binding the Robotic Arm#

Checking Device Recognition#

Make sure the system can recognize the robotic arm. Run the following command to inspect CAN interfaces:

ip link | grep can

Information

Recognized normally: If the command output shows CAN interface information such as can0 or can1, the system has recognized the robotic arm correctly. You can proceed with binding.
Device not recognized: If the command returns no CAN interface information, the system has not detected the robotic arm and the driver must be installed.
```
sudo apt update && sudo apt install ./airbot-configure_5.1.6-1_all.deb -y
```
After installation, disconnect and reconnect the data cable, then run ip link | grep can again to inspect CAN.

Binding Command#

Run the following command to bind the robotic arm:

sudo bind_airbot_device can_left

After binding, disconnect and reconnect the data cable, then run ip link | grep can again to inspect CAN.

Rebinding

If you need to bind the device again for any reason, first run the following command to clear the previous binding configuration:

sudo bind_airbot_device rm

Reconnect the device, then run the binding command again.

Querying Camera Devices#

Use the following commands to query device information for USB cameras and RealSense depth cameras, respectively:

# Query USB camera device numbers and configuration information
lerobot-find-cameras opencv

# Query RealSense camera device numbers and configuration information
lerobot-find-cameras realsense

If the following error occurs during the RealSense query:
Error finding RealSense cameras: failed to set power state
Run these commands to configure udev permissions:
# Create a RealSense udev rule
echo 'SUBSYSTEM=="usb", ATTR{idVendor}=="8086", MODE="0666", GROUP="video"' | sudo tee /etc/udev/rules.d/99-realsense.rules

# Reload udev rules
sudo udevadm control --reload-rules
sudo udevadm trigger
After configuration, reconnect the RealSense camera or restart the system for the rule to take effect.

Modifying the Configuration File#

Before collecting data, configure the hardware devices and dataset parameters. The configuration file has four sections: follower-arm settings (robot), leader-arm settings (teleop), collection settings (run), and dataset settings (dataset).

Quick Start

Download the configuration-file template for your hardware combination:

Hardware Configuration Download Link

Single-arm Play config_Play.yaml

PTK config_PTK.yaml

TOK2 config_TOK2.yaml

TOK4 config_TOK4.yaml

Quest3 quest3.yaml

Follower-Arm Settings (`robot`)#

Follower-arm configuration consists of device parameters and camera parameters:

Device Parameter Configuration#

Parameter	Description	Accepted Values	Example
`robot.type`	Follower-arm type	`airbot_play_follower` `airbot_PTK_follower` `airbot_TOK2_follower` `airbot_TOK4_follower`	`airbot_play_follower`
`robot.port`	CAN interface	`can0`, `can1`, etc. (if bound, this must match the binding name)	`can0`
`robot.id`	Identification ID	Using the same value as `type` is recommended	`airbot_play_follower`

Follower-arm type descriptions:
• airbot_play_follower: AIRBOT Play robotic arm + G2 gripper
• airbot_PTK_follower: PTK series
• airbot_TOK2_follower: TOK2 series
• airbot_TOK4_follower: TOK4 series

Camera Parameter Configuration#

Multiple cameras of two types are supported:

Camera Type	Configuration Value	Features
USB camera	`type: opencv`	General-purpose USB interface; captures RGB images only
RealSense depth camera	`type: realsense`	Intel depth camera; supports RGB + depth images

A complete follower-arm configuration is shown below:

href="#__codelineno-13-1">robot: type: airbot_play_follower port: can0 id: airbot_play_follower cameras: # RealSense camera configuration example right_hand_camera: # Camera name; may be set freely type: realsense # Camera category serial_number_or_name: "243322072684" # Camera SN width: 640 height: 480 fps: 30 color_mode: RGB # Color mode use_depth: false # Whether to enable depth images; disabled by default rotation: NO_ROTATION # Whether rotation is required # USB camera configuration example env_cam_A: # Camera name; may be set freely type: opencv # Camera category index_or_path: "/dev/video2" # Camera port number width: 640 height: 480 fps: 25 rotation: NO_ROTATION # Whether rotation is required

Leader-Arm Settings (`teleop`)#

teleop:
  type: airbot_play_with_E2_leader
  port: can1
  id: airbot_play_with_E2_leader

Parameter	Description	Accepted Values	Example
`teleop.type`	Leader-arm type	`airbot_replay` `airbot_play_with_E2_leader` `airbot_PTK_leader` `airbot_TOK2_leader` `airbot_TOK4_leader`	`airbot_play_with_E2_leader`
`teleop.port`	CAN interface	`can0`, `can1`, etc. (if bound, this must match the binding name)	`can0`
`teleop.id`	Identification ID	Using the same value as `type` is recommended	`airbot_play_with_E2_leader`

Leader-arm type descriptions:
• airbot_replay: AIRBOT Replay teaching arm
• airbot_play_with_E2_leader: AIRBOT Play robotic arm + E2 teaching device
• airbot_PTK_leader: PTK series
• airbot_TOK2_leader: TOK2 series
• airbot_TOK4_leader: TOK4 series

Collection Settings (`run`)#

run:
  fps: 30
  display_data: false
  num_episodes: 25
  episode_time_sec: 60
  reset_time_sec: 60
  single_task: "Grab the black cube"

Parameter	Description	Default
`run.fps`	Data-collection frame rate (Hz), which controls the main-loop frequency	`30`
`run.display_data`	Whether to display camera images and joint states in real time	`false`
`run.num_episodes`	Maximum number of collection episodes; once reached, the program stops automatically and saves the dataset	`25`
`run.episode_time_sec`	Maximum duration of one episode (seconds); when it times out, the episode is saved automatically and the next episode begins	`60`
`run.reset_time_sec`	Maximum collection time after a reset (seconds); the same as `episode_time_sec`	`60`
`run.single_task`	Task-description text written to the dataset as metadata	`""`

Dataset Settings (`dataset`)#

dataset:
  repo_id: test_playe2
  video: true
  num_image_writer_processes: 0
  num_image_writer_threads_per_camera: 4
  video_encoding_batch_size: 1
  push_to_hub: false

Parameter	Description	Default
`dataset.repo_id`	Local dataset directory name. Set a new, unique directory name each time the program runs.
`dataset.video`	Whether to save images as video	`true`
`dataset.num_image_writer_processes`	Number of image-writer processes; modification is not recommended	`0`
`dataset.num_image_writer_threads_per_camera`	Number of writer threads for each camera; modification is not recommended	`4`
`dataset.video_encoding_batch_size`	Batch size for video encoding; modification is not recommended	`1`
`dataset.push_to_hub`	Whether to push the collected dataset to the LeRobot website	`false`

Data-Collection Workflow#

Data-Collection Methods

Two collection methods are currently supported:
- Robotic arm teleoperation: Use a leader arm to control the follower arm in completing a task. This is suitable for conventional desktop manipulation scenarios.
- VR headset control: Wear a VR device for immersive operation. This is suitable for spatial interaction or complex three-dimensional manipulation tasks.

Method 1: Robotic Arm Teleoperation Collection#

Starting the Collection Program#

# Enter the Conda environment
conda activate lerobot
python3 -m lerobot_play.record --yaml <path to your config yaml>

Performing the Task#

Operate the leader arm manually: Use the leader arm (teleop) to control the follower arm in completing the specified task.
Start collection: Press the Spacebar to start collecting data.
Save one episode: After each task is complete, press → (Right Arrow) once to save the current data.

⚠️ Caution: After pressing the key, wait for a period of time while the data is written to disk. The robotic arm is uncontrolled during this period. Do not put down the leader-arm AIRBOT Replay! Collection can continue after the data has been written.
Reset the current episode: If a mistake occurs during a demonstration, press ← (Left Arrow) once to discard the current episode's data and restart the same episode (the episode number does not change).

Exit early: If you have completed the required number of episodes or need to interrupt collection, press Esc to exit immediately. The system saves all successfully completed episodes.

💡 Keyboard Shortcut Reference

Key Function

Spacebar Start collecting data for the current episode

→ Save the current episode and prepare for the next one

← Discard the current episode and restart it

Esc Terminate the entire collection task early

Keys Not Responding?

LeRobot keyboard monitoring depends on X11. If you use Wayland (the default on Ubuntu 22.04+), key presses may not work.

# Edit the GDM configuration
sudo nano /etc/gdm3/custom.conf

# Find and uncomment the following line, setting it to false
WaylandEnable=false

# Save, then restart or sign out and sign back in
sudo reboot
echo $XDG_SESSION_TYPE
# Normal output: x11

Resetting the Environment#

Environment reset: Before each episode, manually restore the experimental environment to its initial state, including:
- Returning objects to their starting positions
- Resetting the gripper to its initial state
- Arranging objects neatly within the camera's field of view
Waiting time: The system waits for the environment to be reset according to the reset_time_sec parameter in the configuration file. Perform the preparation work during this period.

Data Storage Location#

Default storage path
~/.cache/huggingface/lerobot/<repo_id>/

Here, <repo_id> is determined by the dataset.repo_id field in the configuration file.

⚠️ Important: You must change dataset.repo_id for every new collection run. Otherwise, a dataset with the same name will be overwritten and historical data will be lost.

Method 2: VR Headset Control#

Before You Begin
Make sure you are familiar with basic Quest3 operations such as menu navigation and controller use before performing the following steps.

Preparation#

Network connection: Connect the computer and Quest3 headset to the same router with Ethernet cables, ensuring that they are on the same LAN.
Start the VRControl application:
1. Open the VRControl application on the Quest in the specified order.
2. After the application starts, select the Next button in the interface.

Starting the Collection Program#

Run the following commands in a terminal to start data collection (replace the path with the actual configuration-file path):

# Activate the Conda environment
conda activate lerobot

# Start the collection script
python3 -m lerobot_play.record --yaml <path/to/your/config.yaml>

Configuring the IP Connection#

In the Quest VRControl application, enter the computer's wired IP address.

Run the following command in a computer terminal to view the wired-network IP address:

ifconfig
# Locate the inet address for the wired network adapter (such as eth0 or enpXsX)

Verifying the Connection#

After the following information appears, you can use the VR controllers for data collection.

Controller button functions are described below:

Figure 1: Left controller button layout

Figure 2: Right controller button layout

Button	Function
A button (press once)	Start collecting data for the current episode
X button (press once)	End the current episode and proceed to the next one
Y button (press once)	Discard the current episode's data
B button (press once)	End collection and exit
Index-finger trigger (press and hold independently on either side)	Enable control of the robotic arm on the corresponding side
Middle-finger trigger (independent on either side)	Press: close the gripper Release: open the gripper

Caution

Robotic arm control state: During data collection, keep the index-finger trigger on the corresponding side pressed to ensure that the follower arm on that side responds to VR control normally.

State retention: If you release the index-finger trigger during operation, the robotic arm stops responding. When you press it again, the robotic arm remains in the position where it was when the trigger was released.

Information

If you encounter the error error while attempting to bind on address('0.0.0.0',8080):address already in use:

# First query the PID of the process occupying the port
sudo lsof -i tcp:8080

# Then terminate that process
kill -9 <PID>
# Replace <PID> with the actual process PID

Data Storage Location#

Default storage path
~/.cache/huggingface/lerobot/<repo_id>/

Here, <repo_id> is determined by the dataset.repo_id field in the configuration file.

⚠️ Important: You must change dataset.repo_id for every new collection run. Otherwise, a dataset with the same name will be overwritten and historical data will be lost.

Dataset Post-Processing#

After collection is complete, use LeRobot tools to modify and process the dataset.

Supported Operations#

Delete specified episodes: Remove low-quality demonstration data
Split a dataset: Divide training, validation, and test sets by proportion
Merge datasets: Combine multiple related datasets
Manage features: Add or delete data fields

Usage#

For specific procedures and API documentation, see the official LeRobot documentation.

Tip: Back up the original dataset before modifying it.

Data Visualization#

After data collection is complete, use the rerun tool to visualize and validate the dataset, checking collection quality, task execution, and data integrity.

Visualization Command Example#

python3 -m lerobot_play.utils.rerun_vis \
    --repo-id test_playe2 \
    --root ~/.cache/huggingface/lerobot/test_playe2 \
    --robot-type airbot_play_follower \
    --episode-index 0 \
    --max-episodes 1 \
    --batch-size 32 \
    --num-workers 0

Parameter Descriptions#

Parameter	Description	Default
`--repo-id`	Dataset name (the same as `repo_id` in the collection configuration)	-
`--root`	Dataset storage path	~/.cache/huggingface/lerobot/`
`--robot-type`	Follower-arm type (must match `robot.type` used during collection)	-
`--episode-index`	Index of the episode to visualize (starting from 0)	`0`
`--max-episodes`	Maximum number of episodes to load	`1`
`--batch-size`	Data-loading batch size	`32`
`--num-workers`	Number of data-loading subprocesses	`0`

💡 Usage Tips
The dataset may briefly pause when loaded for the first time. This is normal because video must be decoded or HDF5 files read.
After loading is complete, you can play it smoothly, drag the timeline, and inspect multiple camera views and joint trajectories.

Data Replay#

After data collection, replay the motion from a selected episode on the physical robot to validate data quality or demonstrate the task.

Replay Command for a Single-Arm Play Device#

python3 -m lerobot_play.replay \
    --dataset.repo_id test_playe2 \
    --dataset.root ~/.cache/huggingface/lerobot/test_playe2 \
    --robot.type=airbot_play_follower \
    --robot.port=can0 \
    --robot.id=airbot_play \
    --episode_index=0 \
    --fps=30

Replay Command for a Dual-Arm Device (Such as PTK)#

python -m lerobot_play.replay \
    --dataset.repo_id test_03 \
    --dataset.root /home/air/.cache/huggingface/lerobot/test_03 \
    --robot.type=airbot_PTK_follower \
    --robot.left_arm_port=can0 \
    --robot.right_arm_port=can1 \
    --robot.id=PTK_follower \
    --episode_index=0 \
    --fps=30

Caution

Single-arm devices use --robot.port.

Dual-arm devices such as PTK require --robot.left_arm_port and --robot.right_arm_port to be specified separately.

All robot-related parameters (type, id, and port) must exactly match the configuration used during collection. Otherwise, data parsing may fail or control may be abnormal.

Parameter	Description
`--dataset.repo_id`	Dataset name; must match `dataset.repo_id` in the collection configuration
`--dataset.root`	Full dataset path (normally `~/.cache/huggingface/lerobot/<repo_id>`)
`--robot.type`	Follower-arm type, such as `airbot_play_follower` or `airbot_PTK_follower`
`--robot.port`	CAN interface name for a single arm, such as `can0`
`--robot.left_arm_port` / `--robot.right_arm_port`	CAN interfaces corresponding to the left and right arms of a dual-arm device
`--robot.id`	Robot identification ID; using the same value as during collection is recommended
`--episode_index`	Index of the episode to replay (starting from `0`)
`--fps`	Replay frame rate (Hz); using the same value as `run.fps` during collection is recommended to ensure accurate motion timing

Performance Tests#

The following performance-test data comes from benchmark tests of the official LeRobot data-collection kit on different hardware platforms. The results reflect the performance characteristics of the LeRobot kit itself; this adaptation did not modify its underlying collection logic or performance.

Platform	Combination	Mean CPU Load (Load Average)	Save Time (30s Collection) /s	Actual Frame Rate (FPS)
Legion Y7000	`replay+play+1usb`	1.47	2	30
	`play+play+1realsense`	1.47	2	30
	`2replay+2play+3usb`	5.84	14	25
	`2play+2play+3realsense`	7.73	15 17 (depth)	30
AGX	`replay+play+1usb`	2.92	13	30
	`play+play+1realsense`	4.66	14 18 (depth)	30
	`2replay+2play+2realsense+1usb`	7.01	23.1 24 (depth)	29.8
	`2play+2play+3realsense`	9.99	25 26 (depth)	24.4 23.13 (depth)
NX	`replay+play+1usb`	3.09	24	29.6
	`play+play+1realsense`	6.52	33.44 36.1 (depth)	29.6
	`play+play+1realsense+1usb`	8.21	49	25

Architecture	Ubuntu 22.04 LTS	Ubuntu 24.04 LTS
x86_64	✅ Supported	✅ Supported
ARM64	✅ Supported	❌ Not supported

Hardware Configuration	Download Link
Single-arm Play	config_Play.yaml
PTK	config_PTK.yaml
TOK2	config_TOK2.yaml
TOK4	config_TOK4.yaml
Quest3	quest3.yaml

Key	Function
Spacebar	Start collecting data for the current episode
→	Save the current episode and prepare for the next one
←	Discard the current episode and restart it
Esc	Terminate the entire collection task early

LeRobot Play Data Collection#

Environment Configuration#

Software Installation#

System Configuration and Python Version Matrix#

Installation Steps#

Hardware Configuration#

Hardware List#

Binding the Robotic Arm#

Checking Device Recognition#

Binding Command#

Querying Camera Devices#

Modifying the Configuration File#

Follower-Arm Settings (robot)#

Device Parameter Configuration#

Camera Parameter Configuration#

Leader-Arm Settings (teleop)#

Collection Settings (run)#

Dataset Settings (dataset)#

Data-Collection Workflow#

Method 1: Robotic Arm Teleoperation Collection#

Starting the Collection Program#

Performing the Task#

Resetting the Environment#

Data Storage Location#

Method 2: VR Headset Control#

Preparation#

Starting the Collection Program#

Configuring the IP Connection#

Verifying the Connection#

Data Storage Location#

Dataset Post-Processing#

Supported Operations#

Usage#

Data Visualization#

Visualization Command Example#

Parameter Descriptions#

Data Replay#

Replay Command for a Single-Arm Play Device#

Replay Command for a Dual-Arm Device (Such as PTK)#

Performance Tests#

Follower-Arm Settings (`robot`)#

Leader-Arm Settings (`teleop`)#

Collection Settings (`run`)#

Dataset Settings (`dataset`)#