Unitree G1 Demonstration Recording#
The Isaac ROS recorder captures physical Unitree G1 teleoperation demonstrations as MCAP sessions for conversion to LeRobot.
Prerequisites#
Complete real-hardware teleoperation on Thor, connected externally or as a backpack.
Keep CloudXR and the teleoperation launch stack running in their terminals.
Use a separate Thor terminal for the camera checks and recorder.
Review Recording Quality before the first take.
Warning
Use at least two people for recording:
One person for teleoperation
One person to monitor the robot for safety purposes
Workspace Prep#
Before recording, recreate the physical workspace setup.
Keep the table, object appearance, robot pelvis position, and head pose consistent across the dataset.
On the Thor host, outside the container, confirm the kernel sees the RealSense camera over USB:
lsusb | grep -i intel
In the same terminal, enter Isaac ROS:
isaac-ros activateInside the container, confirm the camera is visible:
rs-enumerate-devices
The output should include a connected RealSense with its serial number and firmware version.
Confirm the ego-view camera frame in RViz:
rviz2
Select Add > By topic > /realsense_d435_rgb/color/image_raw/Image > OK and verify the workspace is visible before starting the recorder.
Close RViz to return to the shell before starting the recorder.
Start the Recorder#
In the Isaac ROS terminal used for the camera checks, run the recorder:
ros2 run isaac_ros_unitree_g1_recorder record -- \ task_description:="move the apple to the plate"
Verify that the camera is capturing data. Two methods to check:
Use RViz to view camera topics
Use a second ROS-sourced terminal to list topics with
ros2 topic list
If camera data is not being recorded
Follow the ordered stop procedure before inspecting cables. Once the robot is stopped:
Stop the recorder.
Disconnect and reconnect the RealSense USB cable.
Repeat the camera checks, teleoperation startup, and recorder launch.
Use the terminal UI:
Press Space to start an episode.
Press Space again to save.
Press c to discard.
Recordings are stored in ${ISAAC_ROS_WS}/recordings/.
Recording Quality#
For high-quality datasets, teleoperate slow movements and keep the robot body fixed.
After completing the task, move the left hand out of view and hold it still for 2 to 3 seconds before saving the recording.
For the apple-to-plate task, only the left hand should move and perform the task. Favor side grasps with a bottle-holding-style hand pose so the ego-view camera can see the hand and apple clearly.
Vary the apple and plate position and orientation between episodes, but keep both reachable while the robot stands still. Keep apple type, plate type, table height, lighting, and camera setup fixed for the baseline dataset.
Keep the right arm mostly outside the camera field of view. It is acceptable for right-hand fingers to appear occasionally because the right hand may enter view during deployment.
Keep the baseline dataset focused on clean successful episodes. Add failed-grasp recovery examples only when you intentionally want a broader dataset, and keep them limited so they do not dominate the training signal. Collect at least 200 successful episodes for the baseline training recipe.
Next Stage#
Save or discard the active take, then exit the recorder. Follow
Stop Teleoperation when finished with the
robot. Confirm that the saved session directories exist under
${ISAAC_ROS_WS}/recordings/ and retain them as the source data.
Continue to Unitree G1 Dataset Conversion. If conversion runs on a workstation, copy the complete session directories there first.
See also
Refer to the isaac_ros_unitree_g1_recorder documentation for full keyboard reference, launch arguments, recorded topics, services, build-from-source instructions, and troubleshooting.