Manually control simulation step?

[gkahn13]

Hi @kaihuchen! Great work here. I have a question about Airsim's simulation step that I thought you might know (cross posted on AirSim here).

Is it possible to manually control when and for how long the simulator steps? Basically I wanted to use AirSim with a Gym-like interface (which you've done), but it wasn't clear to me how to do a simulator step. Are you stepping AirSim manually?

[kaihuchen]

In my code envs/airsim/airsimcarenv.py you will find a _step function, which is called by OpenAI Baselines. This _step function is the main connecting point between the OpenAi Baselines and AirSim.

Inside this _step function the state of the AirSim environment is fetched and processed into what I wanted Baselines to see. If you want to control the pace of the simulation step, then this is the place to do it. For my experiments here I did not make any attempt to control the pace and just let it run as fast as possible. As a result the _step function is called 2-4 times per second (Intel Core i7 w 16GB RAM, GPU is nVidia GTX 1070 with 8GB) in my experiments.

[gkahn13]

But does calling _step correspond to a fixed, consistent dt step in AirSim? Or does the step in AirSim depend on your computer specs, how long it takes to compute your desired action, the sim environment, etc?

[kaihuchen]

AirSim runs its simulation in real time (assuming you have reasonable hardware) as a separate process, so when _step queries AirSim's state it just gets whatever AirSim has at the moment, it is unrelated to how fast _step does its computation.

Separately there is a discussion thread about how to run AirSim at faster than real time here: https://github.com/Microsoft/AirSim/issues/522

[gkahn13]

Hm I see. It's unfortunate since debugging deep RL algorithms is a lot easier when you can control the environment step. Hopefully AirSim will expose the SteppableClock interface at some point.

Thanks for your help!