Evolutionary algorithm in Isaac Sim

Hi everyone,

I’m evaluating Isaac Sim for use in evolutionary robotics, and I’d appreciate input from users experienced with large-scale simulation.

Currently, I’m using Gazebo Harmonic to simulate TurtleBot 4, but I run into performance issues when simulating more than a few robots. My experiments require running thousands of short simulations, potentially involving multiple robots per simulation. So I’m looking for a simulator that can:
• Run significantly faster than real time (high Real Time Factor)
• Be automated and scripted for batch execution
• Scale well with the number of robots in the scene

Does Isaac Sim support this kind of workload effectively? Specifically:
• How does the simulation performance scale as I increase the number of mobile robots?
• What are typical Real Time Factors achievable when simulating, say, 5, 10, or 20 robots with basic sensors and locomotion?
• Are there best practices or resources for optimizing performance and automating simulation runs?

Thanks in advance for any feedback or pointers!

This is my first post on the forum. Sorry in advance if I posted in the wrong place. If that’s the case, please point me to the right one.

Hi @paoloelle01! Welcome to the community.

• How does the simulation performance scale as I increase the number of mobile robots?

Please check out Isaac Sim benchmark.

• What are typical Real Time Factors achievable when simulating, say, 5, 10, or 20 robots with basic sensors and locomotion?

If the simulation is running via isaac-sim.sh, then the physics step will adjust itself as best it can to remain real time. If you are running via python.sh, it will run as fast as possible. What real time factors are achievable highly depends on the complexity of your scene.

• Are there best practices or resources for optimizing performance and automating simulation runs?

Please check out Isaac Sim performance optimization handbook

Hi @zhengwang,

Thanks for the information.

Are there any example scenarios that illustrate how environmental complexity can affect simulation speed?

In my case, the scenario will be simple: just a ground with walls, some cylindrical objects, and one or more TurtleBot 4 robots. My main concern is that the LiDAR and infrared sensors rendering might slow down the simulation, at least based on my experience with Gazebo.