I’m encountering a performance issue in Isaac Sim when using RTX Lidar and Camera sensors on vehicles. The FPS drops significantly with an increase in the number of vehicles.
Configuration:
Scene: Two vehicles, each equipped with an RTX Lidar sensor and a Camera sensor.
Desktop Specs:
CPU: i7-11700
GPU: A6000
RAM: 128G
OS: Ubuntu 20.04.6
Performance Observed:
One vehicle: 25~ FPS
Two vehicles: Below 10 FPS
Steps to Reproduce:
Import a vehicle model to Isaac Sim.
Utilize the Vehicle Wizard for vehicle dynamics setup. Reference
Add a ground plane and place some random items on it.
Establish RTX Lidar and Camera sensors via omnigraph.
Attempts to Resolve: I’ve spent weeks troubleshooting this, experimenting with various lidar configurations and settings tweaks, but haven’t found a solution yet. Tried to track down the issue with profiler and htop but have no idea to go further.
Sorry your having trouble. I would need more details about your scene, render output size of the cameras, and lidar configurations.
With one vehicle at 25fps on an A6000, it seems like this is already a rather heavy scene?
Generally, if it ran 25fps with render one camera, and assuming rendering is the bulk of your time, it would run ~13 fps with another identical camera in the mix.
There are a few things you could troubleshoot. What is the scene like without any camera or lidar? Are you using the debug viewer for the lidar point cloud? (don’t, it’s slow). What happens if you don’t output to ROS? What if you just use the lidar? what if you just use the cameras?
Without having a scene to trouble shoot, I don’t know that I can help more.
The good news is, for our next release, we have made progress on making multiple camera/sensor and ROS output faster. Mostly it’s not a matter of actually making the scene render faster, but of making sure we do async rendering, and that data collection and sending to ros is optimized.
Hopefully, with the next release this month, you can try your scene again and have better luck.
The camera render settings and Lidar configurations are set to default: the camera resolution is 1280x720 and the Lidar is configured as example_rotary. I tried decreasing the camera resolution to 640x480 and adjusted the Lidar render product to the same dimensions. I also changed the Lidar configuration to use fewer rays (example_rotary 16). However, these changes did not significantly increase the FPS; it only improved by approximately 2 FPS. Additionally, I noticed that the simulation is not utilizing the full capacity of the GPU and CPU. Could there be some Isaac Sim settings that I haven’t configured correctly?
Based on your suggestions, I conducted several tests:
Running just the cameras for both vehicles with the ROS application open yielded around 38 FPS; this dropped to 25 FPS when the ROS application was running.
Using only the Lidar for both vehicles with the ROS application turned off resulted in around 50 FPS, which decreased to 40 FPS when the ROS application was active.
A scene without any cameras or Lidar ran at approximately 80 FPS.
I’m also curious to know whether grouping all sensors into a single action graph or separating them into individual action graphs would improve performance. I’ve created a simplified scene with settings similar to those in my project, except the vehicle models are different. Could you assist me with this?
I don’t think keeping things in a single graph matters. But you could also check that by using the omni.kit.profiler.tracy extension by hitting F5 to start/stop the profiler.
