Question about object mass realism: low inferred masses from asset densities

[kiwi142857]

Hi, thanks for maintaining LIBERO!

I inspected several object assets in upstream LIBERO at commit 8f1084e3132a39270c3a13ebe37270a43ece2a01 and noticed that their collision geoms consistently use density="100", without an explicit body <inertial> element.

Under the standard robosuite configuration (inertiagrouprange="0 0"), this appears to produce very low object masses:

Object	Collision geoms	Inferred mass
akita_black_bowl	40 boxes	5.59 g
porcelain_mug	22 boxes	16.46 g
glazed_rim_porcelain_ramekin	25 boxes	9.66 g
milk	4 boxes	31.67 g
butter	1 box	5.25 g

These are static calculations from the XML, not measurements of model.body_mass in a running LIBERO environment. For each box, I used:

mass = density × 8 × size_x × size_y × size_z

Here, MuJoCo's box size values are half-extents. I summed the masses of the group-0 collision geoms and excluded the group-1 visual mesh.

Examples

• [Porcelain mug asset](https://github.com/Lifelong-Robot-Learning/LIBERO/blob/8f1084e3132a39270c3a13ebe37270a43ece2a01/libero/libero/assets/turbosquid_objects/porcelain_mug/porcelain_mug.xml): the inferred mass is approximately 16.46 g, which seems unusually low for a ceramic mug.
• [Milk asset](https://github.com/Lifelong-Robot-Learning/LIBERO/blob/8f1084e3132a39270c3a13ebe37270a43ece2a01/libero/libero/assets/stable_hope_objects/milk/milk.xml): the main collision box measures approximately 5.25 × 5.25 × 10.95 cm, or 302 mL. If this represents a filled carton, a total mass of approximately 31.67 g seems too low. An empty carton would be a different interpretation.
• [Butter asset](https://github.com/Lifelong-Robot-Learning/LIBERO/blob/8f1084e3132a39270c3a13ebe37270a43ece2a01/libero/libero/assets/stable_hope_objects/butter/butter.xml): the collision box occupies approximately 52.47 cm³ but has an inferred mass of only 5.25 g.

Questions

1. Is this low density an intentional choice for numerical stability or easier manipulation, rather than physical realism?
2. Does any environment-loading step override these densities or inferred masses?
3. Are these objects intended to represent empty packaging or filled products?
4. Would it be useful to document the effective masses and, optionally, provide a physically calibrated asset configuration?

I understand that changing the default dynamics could affect compatibility with existing demonstrations and checkpoints. An opt-in configuration could preserve the original benchmark.

I have not yet tested the effect on policy success rates, so this issue is a request to clarify the intended physical parameters rather than a claim about benchmark performance.

Thank you!