I am using a Nvidia Quadro K6000 GPU (12GB) with Ubuntu 16.04 and python 2.7.12. I am having a problem in achieving high FPS (the link given to the github issue was answered by Erwin in which he achieved 280 FPS with GTX 1080) with my computer. I have installed pybullet using pip and the attached code (shared in the github issue as well) can be used to measure the average FPS.
I am attaching the text file output of the code. I am getting a 30 FPS output when averaged over 100 iterations, which is almost the same as I am getting on a non-GPU system. In addition, I am attaching the .json performance logger file which shows the time taken for individual processes (can be viewed in chrome browser using about://tracing) while running this code. Here, as also mentioned in the github issue #1545 rendering the image only takes microseconds (260 microseconds in my case) while most of the time is spent in waiting for the image (from #1545: "Note that very little time is currently spend in rendering (less than 200 microseconds on my 1080), it is mainly copying data / waiting that makes is several milliseconds."). I am attaching a screenshot as well for further reference.
I am planning to shift my Deep RL simulation from gazebo to pybullet as the latter is much more flexible and coder friendly
. I am also interested to know if pybullet is not the correct SDK for using bullet physics engine for RL research is there any python/C++ based tool which has a well established GPU support system? Any help at all would be appreciated thanks!!!