I am facing a weird error. Whenever I call gym.create_camera_sensor(env, camera_props), I get an error: [Error] [carb.gym.plugin] cudaImportExternalMemory failed on rgbImage buffer with error 999.
I am trying to get the RGB images from unitree a1 environment in this repo: (GitHub - leggedrobotics/legged_gym: Isaac Gym Environments for Legged Robots).
Even if I try to run the interop_torch.py example then also the same error comes up.
I encountered the same problem with RTX3070, Ubuntu 20.04, did you solve it?
If you are getting the same error while running the interop_torch.py then delete your conda installation try installing conda with py3.7 as the base and then try to install the issacgym environment (which I did and it worked, as my initial conda base env was with python3.8). By the way, issacgym recommends ubuntu 18.04 so if this solution does not work then you might have to install ubuntu 18.04. For me this error was generated in pop-os 20.04 with conda base as python-3.8.
did you solve it by changing python version? I am using python3.7 on ubuntu 22.04 and have the same error
I use laptop with both integrated card and a rtx 3070, and thus it usually runs in hybrid mode. In this case, this error will be raised. But after disabling the integrated one and using the discrete card only, this error can be fixed.
I think you can do this in bios setting. Hope it can help.
Seems like there is no such option in the bios setting of my PC. Did you switch it using the bios? I also tried nvidia-prime via sudo prime-select nvidia. Moreover, I am using Preview Release 4, are you also using this version?
I solve the issue by
export VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json
Thanks! You saved me! I’m using Isaac Gym Preview Release 3 in Ubuntu 24.04 and with a RTX 4060 with Python 3.7