How to debug when `gym.simulate()` hangs?

Hi,

I am having some issues with running my model in ISAAC gym. But after a few simulation steps, gym.simulate() will freeze/hang. And the entire code will stop there, and it does not give any error. I wonder how to debug in this case? And what would people suggest doing in this case? Or are there some hypotheses about when this would happen? Thanks.

1 Like

Hi @taocc,

One thing you can do in this case is record a PhysX debug log, by setting the GYM_PVD_FILE environment variable:

export GYM_PVD_FILE=testpvd

Alternatively, you could point our team toward a reproducible test case.

You can also try using --device=CPU command line parameter to determine if the issue is related to either running in GPU PhysX, or using the tensor communications path. The --device=CPU parameter forces the tensor communications path off, which will also disable the GPU PhysX path by default.

You can also use --device=CPU along with --physx_gpu to force GPU PhysX simulation on while still disabling the tensor path - this mode copies GPU results back to the CPU and returns CPU tensors for any tensor APIs you use.

Take care,
-Gav

I was having a similar problem with gym.simulate(), for me it would always occur at the same time at end of an episode when the simulation would call reset(). I found that the issue related to gym.set_dof_state_tensor_indexed()which was called during the reset and applyed in the next simulate call. Commenting out this part of the code stopped the hanging, but I am not sure why it was not working as I have implemented the function in the same way as FrankaCabinet.py

I hope that can save others from experiencing similar headaches as I had for the past few days.

1 Like