How to use multiple process calculate controller?

[Wan-ZL]

Hi,

I'm using dozens of drones in simulation, and the "computeControlFromState" takes a lot of time. I tried the Process from multiprocessing but got errors.

So, I would like to know how to make the for loop below run in parallel and speed up the code.

ctrl = [DSLPIDControl(drone_model=DroneModel.CF2X) for i in range(num_drones)] # Initialize "num_drones" controllers

for i in range(num_drones): # Compute control for each drone

action[str(i)], _, _ = ctrl[i].computeControlFromState(control_timestep=env.TIMESTEP, state=obs[str(i)]["state"], target_pos=TARGET_POS)

Thank you.

[JacopoPan]

Hello @Wan-ZL,

Process or binding a C implementation would be my first guesses to try and speed-up independent computational blocks. Do you perhaps want to fork the repo, open a PR with your multiprocessing implementation and we can see where the errors are coming from?

[Wan-ZL]

Hi @JacopoPan,

I forked the repository and here is the link: https://github.com/Wan-ZL/gym-pybullet-drones

Under the examples folder, I created an example called 'fly_parallel.py' which is the copy of the 'fly.py' but replace the single process PID control part (line 177 in fly_parallel.py) to the multi-process PID control (line 164 in fly_parallel.py).

The code doesn't raise any error but PyBullet GUI is frozen and the pybullet keeps printing the "pybullet build time: Jan 16 2022 01:49:51" as shown below

So, how to solve this issue? Or do you have other way to make PID control function (action[str(i)], _, _ = ctrl[i].computeControlFromState()) run in parallel?

Thank you!

[JacopoPan]

Hi @Wan-ZL,

sorry for the delay but I was busy with a few other tasks, the multiple prints of pybullet build .. I suspect come from the different processes importing pybullet on their own (as it is an import in class `DSLPIDControl but really only for a couple of Euler/quaternion operations at the very beginning and very end: they could could be factored out as well)

My suggestion is to take the following 2 steps:

• debug your (multiprocess, not multi-threaded) parallelism replacing the control method with a dummy function
• re-factor DSLPIDControl.computeControl() as purely functional (also returning and re-accepting its internal state)