Python bindings to enable standalone physics for RL

[jrtcppv]

I am working on a reinforcement learning application in which I only need the kinematics during training and will use the rest of AirSim (rendering, Unreal, etc) during evaluation/demonstration of the trained agents. To maximize the speed of the simulation, I was considering writing Python bindings directly on top of the C++ code using something like Boost::Python, such that something like the example here could be achieved in Python, specifically in an OpenAI Gym environment. I understand from this issue that the normal python module is not available when Unreal is disabled.

My questions are:

1. Is this a good approach for maximizing execution speed given my need for kinematics only?
2. Has this already been done?

Many thanks for any feedback!

[jonyMarino]

Hi @jrtcppv!
You can run AirSim in headless mode. Take a look at this comment: #1224 (comment)
To do a reproduction, it will be easier when we have the ExternalPhysicsEngine ready.

Keep this discussion updated with your progress.
Good luck!

[jrtcppv]

Thank you for the response! I might try it out to compare with running the headless server. If I come up with anything I will post benchmarks here.