[Question] Can I teleoperate Isaac Lab (imitation demonstrations collection) by publishing my own joint commands over ROS topics?

[DaeyunJang]

Environment

Isaac Lab 2.1.0 (also 2.3.0)

Post (short)

Hi — I have a question about using ROS/ROS2 to teleoperate robots inside Isaac Lab for imitation-learning dataset collection.
I followed tutorial to collect demonstration data for Isaac Lab imitation learning by teleoperating 'franka pandas' robots (using keyboard).
https://isaac-sim.github.io/IsaacLab/main/source/overview/imitation-learning/teleop_imitation.html

isaaclab.teleop.mp4

What I want to do

I have custom control logic (not keyboard/joystick) that produces joint commands for a robot. I want to publish those commands as ROS topics and drive a robot inside Isaac Lab (so I can record demonstrations for imitation learning).

My questions

1. Is it supported / recommended to control Isaac Lab robots by publishing joint commands over ROS/ROS2 topics? (e.g. sensor_msgs/JointState). If yes — which topic/message types are commonly used/supported for low-latency control in Isaac Lab?

2. If it’s possible to implement this with ROS, would there be any response delay? In the video the control wasn’t via ROS, and it seemed more responsive. (Like the Franka Panda robot arm in the tutorial.)

3. Are there example setups (minimal code snippets) that publish topic of joint commands inside Isaac Lab?

Any advice would be greatly appreciated.

Thank you.

[RandomOakForest]

Thank you for posting this. Yes, it is possible to teleoperate robots in Isaac Lab using ROS or ROS 2 by publishing joint commands, and this approach is supported for custom demonstration collection workflows. Isaac Lab supports ROS/ROS2 integration, allowing external control via standard topics. However, direct Python control (like keyboard teleop in the Isaac Lab tutorial) is typically more responsive compared to ROS message publishing due to network, serialization, and bridge overhead.

I'll move this to our Discussions section for follow up. In the meantime, here is a summary to consider.

ROS Topics for Isaac Lab

• sensor_msgs/JointState: This is the most commonly used message type for transmitting joint positions, velocities, and efforts for robot arms in ROS. Isaac Lab supports subscribing to JointState for joint commands, but for low-latency control, you may also consider using:
• trajectory_msgs/JointTrajectory: Useful if publishing sequences or time-stamped joint positions. For real-time teleop, JointState is simpler and more commonly adopted.
• Custom Command Topics: Some Isaac Lab ROS integration variants allow topics like /joint_command or /joint_states, depending on the robot configuration. Review the latest Isaac Lab ROS bridge documentation to confirm the exact topic names.

Response Delay Considerations

• Direct Python Control (keyboard, joystick): Lowest latency — the command signals are sent directly within the simulation runtime, yielding near-instantaneous response, as observed in the Isaac Lab teleop tutorial.
• ROS/ROS2 Control: Slightly higher latency due to message serialization, publishing/subscribing, and bridge conversion. Practical delay depends on simulation rate, publish rate, network setup, and bridge efficiency. For most demonstration collection (including imitation learning), the delay is small (typically on the order of tens of milliseconds), but not always negligible for very high-frequency fine control. For low-latency requirements, set high ROS publish rates and run the ROS bridge on the same machine as Isaac Lab.

Example Setup: Publishing Joint Commands

Below is a minimal ROS Python publisher example for real-time joint commands to Isaac Lab's Franka robot:

import rospy
from sensor_msgs.msg import JointState

rospy.init_node('franka_joint_command_publisher')
pub = rospy.Publisher('/franka/joint_states', JointState, queue_size=10)

js = JointState()
js.name = ['panda_joint1', 'panda_joint2', 'panda_joint3', 'panda_joint4', 'panda_joint5', 'panda_joint6', 'panda_joint7']

rate = rospy.Rate(100)  # 100 Hz for near real-time
while not rospy.is_shutdown():
    # Replace with your custom control logic
    js.position = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
    pub.publish(js)
    rate.sleep()

• Adjust the topic names to match your Isaac Lab setup (/franka/joint_states or /joint_command as appropriate).

In Isaac Lab, your robot should subscribe to this topic via its ROS bridge configuration. Reference the Isaac Lab ROS bridge documentation for setting up subscribers and confirming topic/callback mapping for your robot model.

Isaac Lab Documentation References

• The Isaac Lab ROS teleoperation documentation confirms support for teleoperation via ROS for imitation learning, though it notes higher precision and responsiveness when done directly via Python tools.
• Review the ROS Integration Tutorial for Isaac Sim for best practices, latency optimization tips, and advanced configuration for your robot.

[DaeyunJang]

At first, thank you for your reply.
I'm gonna follow your guide.
Before doing this, I have some additional questions.

1. Make subscriber?
   It seems that the provided reinforcement learning environment code samples don’t have a ROS system implemented.
   So, it should be correct to follow Isaac Sim’s ROS-related documentation and implement it with the Python API, right?
   The reason I’m asking is that I was able to create a robot subscriber with Omnigraph, and I’m wondering—if I save that as a USD and then just load the robot in Isaac Lab, will the subscriber still work automatically?

2. Are there any specific settings I need to configure in order to call ROS from the API? (For example, Python path, sourcing ROS, etc.)