Open YitianShi opened 5 months ago
As I understood, your environments do not require the multi-GPU setup. Could you try running the code on a single GPU to see if the same error still occurs? In any case, this error seems to be related to Isaac-Sim and TensorAPI instead of Orbit itself. To get better feedback, I would recommend to pose your questions in their channel. I heard from an experience where it solved by switching the GPU model to a 4060
We also contacted the Isaac Sim Team about it a while ago, @Mayankm96 maybe we can ping them again
Question
Hi, I'm setting up a robot bin picking simulation with top-down camera that captures semantic or instance segmentation after each grasp attempt. My simulation always get
CUDA error: an illegal memory access was encountered
when the jacobian of my robot is got from Physx:Other modalities of camera such as rgb, normals and depths are working fine, while only the semantic or instance segmentation will cause such crash.
The similar issue that I found is: https://forums.developer.nvidia.com/t/multiple-isaac-sim-containers-on-one-gpu-fails-with-cuda-illegal-memory-access-in-omni-physx-tensors-plugin/268134
Where I'm also sure that my GPU memory is far enough than exhaustion since I'm using 4 RTX4090GPUs to run only 3 bin-picking environments with only 2 objects in the bin.
The error message looks like:
After setting CUDA_LAUNCH_BLOCKING=1, gives: