Dynamic gripper-width filtering at inference

[carolina1998-tech]

Hi maintainers,

Thank you again for sharing this great work—really impressive and useful!

I’d like to ask whether the current inference pipeline allows the network to filter or recommend gripper widths that are smaller than the mechanical maximum.
For example, my gripper can open up to 85 mm, but for a 30 mm-wide object a 50 mm (or even 30 mm) pre-shaping would be safer and would avoid collisions. Right now it looks as if every grasp is generated with the full 85 mm aperture, so the collision checker discards most candidates.
Is there already a parameter or post-processing flag that lets me clamp the predicted width to the object size, or to a user-defined range (min_w, max_w), before the collision check? If not, could you kindly suggest the file/entry-point where I could inject this width constraint myself?

Thanks a lot for your time and for the excellent codebase!

Best regards,

[adithyamurali]

Hi @carolina1998-tech unfortunately I've been swamped to implement this feature... but will do so when I get a chance.

[adithyamurali]

The idea is essentially to apply some ray casting between the finger tips (in the contact direction) and the object point cloud (may need to convert to mesh using marching cubes). This will tell us where the gripper contacts the object, and one can calculate the closing width