SHARED_MEMORY_SERVER support

[YunfeiBai]

Having an equivalent version of the function below for headless mode, to start an in-process physics server and physics client in process with shared memory, which PyBullet can connect to

bullet3/examples/SharedMemory/SharedMemoryInProcessPhysicsC_API.h

Line 14 in a506879

B3_SHARED_API b3PhysicsClientHandle b3CreateInProcessPhysicsServerAndConnectSharedMemory(int argc, char* argv[]);

[mbusy]

Hi, any updates on the topic ?

[erwincoumans]

pip3 install pybullet -u
python3 -m pybullet_utils.runServer&
Is that what you were looking for?

[mbusy]

Hi Erwin, more or less yes! (without a GUI, but I'm guessing that I can use something like pybullet.connect(SHARED_MEMORY_SERVER) to do that).

I was actually wondering if creating several physics servers that way would allow me to run multiple simulations in parallel (using multiple cores)? Thanks again !

[erwincoumans]

To run many simulations in parallel on many cpu cores, I recommend using pubullet_utils.bullet_client and multiprocessing like this example.

https://github.com/bulletphysics/bullet3/blob/master/examples/pybullet/gym/pybullet_robots/panda/batchsim3_grasp.py

[mbusy]

Thanks!

[erwincoumans]

Sure. One more example using MPI instead with pybullet_client is deep_mimic:
python3 mpi_run.py --arg_file train_humanoid3d_spinkick_args.txt --num_workers 4
mpiexec -n 16 python3 -m pybullet_envs.deep_mimic.DeepMimic_Optimizer

This latter is rather finicky, since Tensorflow and MPI sometimes break their API.