Dataset issues

[czlong24]

During run dexgrasp_ Generation, an error occurred while training GrassIPDF: FileNotFoundError: [Errno 2] No such file or directory: 'data/DFCData/poses/core/browl-530c2d8f55f9e9e8ce364511e87f52b0';FileNotFoundError: [Errno 2] No such file or directory: 'data/DFCData/meshes/core/bowl-d1addad5931dd337713f2e93cbeac35d However, there are no pose or meshes in the DFCData folder in the downloaded dataset. May I ask if there is an error in the code or if the dataset name is incorrect.

[mzhmxzh]

The meshdatav3.tar.gz should be extracted and renamed to meshes, the datasetv4.1.tar.gz should be extracted and renamed to poses. We will clarify this in the README files soon.

[lptl]

May i ask how the pcs_table.npy is generated?

[mzhmxzh]

First, use generate_object_pose.py to randomly drop the object from a high place for 100 times. Nest, use generate_object_table_pc.py to take depth images of the object and table, project them into partial point clouds, then fuse them into a complete point cloud. Furthest point sampling is used to down sample the point cloud to 3000 object points and 1000 table points.

[lptl]

looks like the generate_object_pose.py has some error on my machine, so that the progress never moves. And may i ask which version of sapien the code is using? I am using 1.1.1 right now.

[mzhmxzh]

I fixed a bug in the newest commit on the main branch. How about now? I am also using SAPIEN 1.1.1 Perhaps try running the script with command line argument --n_cpu 1. If it still doesn't work, check whether the Vulkan renderer is properly installed.

[lptl]

it turns out that i can run this successfully without using coard.urdf, ie, using the else branch not the if branch. But for each folder, i have its related .urdf file.

[mzhmxzh]

If that is the case, then I guess it may be because the urdf loader of SAPIEN supports only up to 64 object pieces, and you are trying to load a decomposed object with more than 64 convex pieces. The else branch turns the object into a complete convex hull. I think it doesn't hurt the simulation, since the only collision is between the object and the table plane.

[lptl]

May I ask what kind of system you are using? Or even more exactly, the environment used to run the dexgresp_policy. I am currently using Ubuntu 20.04 and Python 3.7.0 and have the following error: [Error] [carb.gym.plugin] cudaImportExternalMemory failed on rgbImage buffer with error 999

[wkwan7]

May I ask what kind of system you are using? Or even more exactly, the environment used to run the dexgresp_policy. I am currently using Ubuntu 20.04 and Python 3.7.0 and have the following error: [Error] [carb.gym.plugin] cudaImportExternalMemory failed on rgbImage buffer with error 999

Hi, Ubuntu 20.04 and Python 3.7.0 is OK. It seems something is wrong during the rendering of the simulator. I recommend to:

1. First make sure IsaacGym is correctly installed. (Ensure that Isaac Gym works on your system by running one of the examples from the issacgym/python/examples directory, like joint_monkey.py)
2. check Vulkan by vulkaninfo, this is crucial if you want to use the visual input in the simulator.

[lptl]

Thanks. Both vulkaninfo and example can be executed successfully. I am using IsaacGym preview relase 4. Are you also using this version?

[lptl]

Sorry, i have got the information from README.md. I will try preview release 3.

[wkwan7]

Thanks. Both vulkaninfo and example can be executed successfully. I am using IsaacGym preview relase 4. Are you also using this version?

Is this error in vision-based policy training, and is state-based policy running properly? Also if you are using a multi-gpu machine, you can try to adjust the graphics_device_id here, I know in some multi-gpu machine this may need to be set to a particular device.

[lptl]

Thanks. This error happens when running vision-based policy training. And I have only one GPU on my machine. Seems like the error is more about my own environment issue not the code. I also can't run examples/inter_torch.py from isaacgym.

[zihuuuu]

When I follow the steps to configure the dexgrasp_generation environment, training GraspIPDF,an error will be reported showing AttributeError: 'int' object has no attribute 'backward',Please ask how to solve this problem

[lptl]

Thanks. This error happens when running vision-based policy training. And I have only one GPU on my machine. Seems like the error is more about my own environment issue not the code. I also can't run examples/inter_torch.py from isaacgym.

I solved this issue by using export VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/nvidia_icd.json. This command is written inside the troubleshooting of the installation documentation of the isaacgym.

[lptl]

When I follow the steps to configure the dexgrasp_generation environment, training GraspIPDF,an error will be reported showing AttributeError: 'int' object has no attribute 'backward',Please ask how to solve this problem

i changed the code in models/model.py

[mzhmxzh]

Thanks for pointing out this problem and for solving it! It is fixed in our newest commit. :)