Robustness metric for each grasp

[ghost]

Hello,

I am new to the topic dex-net and now I am reading all these codes.

I only find a python file called "generate_gqcnn_dataset.py" for generating the quality metric for each grasp. However I havn't seen any calculation which involves probability calculation for observation model, grasp candidate etc, which is mentioned in the second paper "Dex-Net 2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds adn Analytic Grasp Metrics" in chapter 4.

Please tell me if I have missed something or which file is designed for the robustness function.

Thank you very much!

[jeffmahler]

@YtDong-Tum The grasp quality metrics are precomputed so the code is not in the generate_gqcnn_dataset.py script. I can't explain in full detail but please see this file for a start.

[pyni]

which file?

[visatish]

@pyni sorry the link is broken. It is this file.

[DYT-code]

Hello community,

I have a question related to this topic.

Is the value "robust_ferrari_canny" given in the dataset "DEX-NET 2.0 DATASET" already the final value we use to train the gqcnn network as the probability of grasp succcess? Can we directly get this value with the code "quality.py" since I've seen in this script a function for calculating ferrari_canny_L1.

Besides, the gripper width is already used in generating the dataset. If I just change the gripper width in gqcnn code when I try to predict a grasp but still with a model trained with old gripper width, I think there will be mistakes. Is that right?

Thank you very much in advance!

[jeffmahler]

Question 1: Yes although the values are thresholded at 0.002 during training. You can also compute the quality of grasps on a 3D CAD model using quality.py but you may want to do so through the CLI.

Question 2: That is correct. The gripper width in the gqcnn code is for grasp sampling only, so the network would still be predicting quality for a narrower gripper width.