Real robot experiment

[han-kyung-min]

Hello,

Thank you for releasing your code.
Is there any tutorial on how to deploy and run the system on a real robot?
It appears that the current README only describes how to run the code in the Isaac Sim environment.

I am using a wheeled differential-drive robot (similar to a TurtleBot) with ROS Noetic on Ubuntu 20.04.
I looked at the closed issue page (#63). However, the provided link guides me to LoGoPlanner, which appears to be a different work, so I am a bit unsure where to start for PG navigation with the real robot.

Thank you in advance for your help.

[wzcai99]

Hi, the deployment process of LogoPlanner and NavDP is similar, and the only difference is to start a different navigation policy server here https://github.com/InternRobotics/NavDP/blob/master/baselines/navdp/navdp_server.py.

[han-kyung-min]

Thanks for the reply.
Unfortunately, I am still a bit confused. Did you mean that running LogoPlanner is equivalent to running NavDP?

Or, did you mean that I need to run navdp_server.py, followed by lekiwi_logoplanner_host.py? It seems that this script does not support a ROS environment.

Do you have any other scripts that support running the system in a ROS (Noetic) environment?

[wzcai99]

For the lekiwi deployment, you only need to replace the line

NavDP/baselines/logoplanner/logoplanner_realworld_server.py

Line 30 in 92cbe77

navdp_navigator = LoGoPlanner_Agent(intrinsic,

to become NavDP_Agent like this:

NavDP/baselines/navdp/navdp_server.py

Line 30 in 92cbe77

navdp_navigator = NavDP_Agent(intrinsic,

And keep running the same lekiwi_logoplanner_host.py. For ROS based deployment, please refer InternNav repo: https://github.com/InternRobotics/InternNav?tab=readme-ov-file#-community-deployment--best-practices

[han-kyung-min]

Thanks for the prompt replies.
Sorry for asking similar questions—I think I may have phrased my earlier questions incorrectly.
Could you kindly answer the questions below?

(1) Can I run NavDP on my wheeled robot (not LeKiwi, but closer to a TurtleBot) running ROS1 Noetic?

(2) If so, could you please elaborate on how to do it?

(3) If NavDP doesn’t support ROS1, or if making it work with ROS1 would require significant effort, please let me know. In that case, I will give up on this package.

Best

[wzcai99]

Hi, the core of NavDP runs in a Server-Client architecture. The "Server" (which runs the deep learning policy/LogoPlanner) is independent of ROS and does not care about the specific robot hardware. It only needs observations (RGB-D images) and goals, and returns velocity commands. Therefore, it is fully compatible with a ROS1 Noetic setup.

Step 1: Setup the Server (Policy Side)
You can run the NavDP/LogoPlanner server on any machine (even a powerful PC) that is connected to the same local network as your robot.
Modification: You only need to ensure the IP address and port in the server configuration match what you will use in your client code. No other changes are needed here.

Step 2: Create a ROS1 Client (Robot Side)
On your robot (running ROS1 Noetic), you need to write a custom Python node. This node acts as the "Client".
Functionality: This client should mimic the logic in lekiwi_logoplanner_host.py, but using ROS1 syntax (e.g., rospy).
Data Flow:
Subscribe: Use ROS1 to get RGB-D images and the current goal coordinates.
Send: Package this data and send it to the Server via the API.
Receive: Get the predicted velocity trajectory from the Server.
Publish: Convert the velocity command to a ROS1 message (e.g., geometry_msgs/Twist) and publish it to your robot's base controller.

[han-kyung-min]

That helped me understand your system better. I was expecting ROS-based communication, but, apparently, NavDP uses a different communication method. I will try to write a ROS-based client script as you suggested
.
If possible, please do not close this issue, so that I can ask new questions or reach out again if I get stuck.

Best

[han-kyung-min]

I got the myrobot_client.py working. Yet, the msg I get from the navdp_server.py doesn't look like a standard cmd_vel.

Apparently, navdp_step_xy() returns

return jsonify({
    'trajectory': execute_trajectory.tolist(),
    'all_trajectory': all_trajectory.tolist(),
    'all_values': all_values.tolist()
})

This looks like a trajectory or a next waypoint to follow (i.e., not a velocity command). It seems like those two have different shapes and length.
Could you please explain what trajectory, all_trajectory, all_values represent, and what each of them is used for?