issues with accuracy

[mhyatt000]

several objects floating around, self-collision, etc.
apologies if i making a mistake, but this is the version of the data I see on huggingface

[Vector-Wangel]

Since the pipeline is automatic, it will automatically detect all the objects with the similar semantic (in this case red bowl). And a motion detector will be applied to every objects to identify whether they are static or grasped by hands. The "floating" objects are static (marked by grey) and since we currently don't offer a complete scene reconstruction, they might look like "floating". This will not affect the performance of the target objects.

Similar way with object self collision, we specifically rescaled the object in hand, will later apply this to all the object meshes, though it currently won't affect data quality if the object is fully static.