Faster than real time SITL

#1

Hey everyone!

I’m attempting to utilize PX4/Gazebo to train a ML/AI algo to recognize target waypoints visually. The main concern is it doesn’t appear possible to run this setup in SITL any faster than real-time, which for training a ML algorithm would be very painful. Searching through the previous issues I’ve noticed a couple stale threads regarding this issue. Is anyone aware of any other configurations/software/simulations that might work for this or might have worked around this limitation themselves? The reason I’m making a new topic is to also open discussion on any other issues the community might see with attempting to do this as I’m more experienced on the software/ML/AI side of things rather than the drone side.

Any help is appreciated, this is a fun and fascinating project and I’m hoping it’s successful,
Thanks!

#2

I’m working on this. For progress see:

Currently it works for jMAVSim and I’m now working on support in Gazebo.

#3

Available in upstream and in release 1.9: PX4_SIM_SPEED_FACTOR=10 make px4_sitl gazebo_plane

#4

You can also look at something like reinmav-gym, if you are interested.

This separates learning with emulating the firmware, making the learning process much more efficient