Optimal Grasp is inferred as collision in the Model.

[bk-poudel]

Dear NVLabs Team,

Thank you very much for developing such a wonderful grasp pose detection algorithm. I have successfully run the system and it works very well. However, when I enable collision checking, some grasps that appear optimal are marked in red (collision). Could you kindly clarify if I might be missing something in the configuration or preprocessing steps?

Additionally, I would like to use the model with my Unitree Z1 arm. From your experience, would it be better to retrain the model specifically for the Z1 gripper, or is it sufficient to transform the provided gripper frames to match the Z1 gripper using frame transformations?

Picture for Reference.

[adithyamurali]

Hi @bk-poudel

Yes, it is indeed strange that so many grasps are deemed to be in collision. Could it be that you had set the collision threshold to be too high?

Yes, it would be ideal to train a new model for the Unitree Z1 gripper. I'm assuming it is a revolute 2 fingered gripper. However, if you just retarget the grasps from the existing models, you may be able to get some non-trivial performance already, depending on the object

[bk-poudel]

Thank you for your reply @adithyamurali ,

I have kept collision threshold to be low 0.005 - 0.02 and I am getting same results. Now, I am trying to retrain the model for Z1 arm but I have a question. Do I just need the urdf file for the gripper or do I need to provide the urdf file for the gripper and the Z1 arm so that the whole dynamics could be learnt. Since, the z1 arm is not as flexible as franka. Optimal grasp detected by the graspgen could not be executed by the gripper.