Tilted output grasping pose due to hardcoded grasp depth

[Moreland-cas]

Hello! Thank you for your amazing work on AnyGrasp. It's been really helpful.

I'm currently running AnyGrasp on my custom point cloud data and have encountered a few issues. I've attached the predicted grasp results for my data with the top_down parameter set to False and True, respectively.

1. top_down=False

When top_down=False, AnyGrasp generates a large number of grasps. However, since the hardcoded depth is set to 2cm and the handle in my data is relatively short, the grasps appear to be tilted.

2. top_down=True

When top_down=True, the number of available grasps is significantly reduced. In fact, no grasps are found in the middle drawer.

I'm wondering if my analysis is correct: does the hardcoded depth cause the tilted grasps? Is there a workaround to address this issue? Thank you in advance for your help!

[chenxi-wang]

Hi, when top_down is enabled, the predicted grasps are filtered by the approach direction, which is a subset of that in top_down=False. That's why you couldn't get the exact predictions you want. The predicted depths are in {0.5, 1, 2, 3, 4}cm.

The easiest way is to disable top_down and the predictions, and manually enlarge the predicted grasp depth by 1-2 cm.

Also, you could try cropping the input point cloud to keep only a small region near the middle handle (set dense_grasp=True if necessary), or combining the predicted translations with a fixed rotation.