search

PKU-EPIC / UniDexGrasp

Official code for "UniDexGrasp: Universal Robotic Dexterous Grasping via Learning Diverse Proposal Generation and Goal-Conditioned Policy" (CVPR 2023)
145 stars 13 forks source link

PointNet checkpoint path in the archived files #30

Open ubt0001 opened 5 months ago

ubt0001 commented 5 months ago

Hello,

In issue #25 there is a request for the PointNet checkpoint to facilitate the addition of new objects and a link that leads to multiple archived files was provided. Could you kindly provide some hints on the path to the model checkpoint after extracting the files from that link? The link offers extracted features from different object datasets, but it appears to not include the model weights. Is there something I might have missed?

Thank you in advance for your assistance.

ubt0001 commented 5 months ago

Additionally, this comment points out that the features of many objects have been set to zero. Could you provide an example of training with all zero object features?

wkwan7 commented 5 months ago

Hi @ubt0001! We have not released the pre-trained PointNet model checkpoint. It is very easy to set the input visual feature to zero at this line; simply change "0.1 * self.visual_feat_buf" to 0 here.

ubt0001 commented 5 months ago

When multiple types of objects are loaded in the simulation environment, does directly setting the features of all objects that do not have feature vectors to zero match the experiment settings in the paper?