[Galbot G1] Left gripper fingertips visually penetrate the bread in put_bananas_bread

[CL-52]

Environment

• Local Genie Sim deployment (Isaac Sim / PhysX, headless)
• Robot: Galbot G1 (galbot_one_golf), URDF galbot_one_golf_fixed_base.urdf
• Scene: scenes/scene_galbot_put_bananas_bread/put_bananas_bread.usda
• Task: put the bananas and bread onto the plate
• The clipped object is the round bun (bread), not the orange. Bread is dynamic (mass 0.09 kg). The orange in this scene is kinematic.

What happens

When the left gripper closes on the bread, the black fingertips go into the bun in the left-wrist camera. The visual overlap looks like centimeters. The hand only stops after the palm / knuckle touches the bread.

What we already checked

Finger contact is authored on the gripper instance prototypes. The URDF collision branches are disabled.

1. Bread collision is aligned with the visual mesh. After fitting the bread collider to the visual mesh, only 2 of 2924 visual vertices stick out, by at most about 0.38 mm. The bread shape is not what lets the fingertip sink in.

2. Proximal finger links do contact the bread. base_link / link1 / link2 use convexHull of their own visual mesh. Logged contacts against bread have separation from about -0.23 mm to a few mm. The solver penetration on the bodies that actually collide is sub-millimeter, not centimeters.

3. Distal tips do not stop the bread. The links that visually enter the bun are link3, link3_1, link3_2, link3_3, link3_1_1, link3_2_1.

4. Approximations we tried on those distal links:

   • convexDecomposition / CoACD compound hulls: a large fraction of visual vertices sit just outside the hulls (about 0.8–1.3 mm), and there are gaps between hulls. The bun can enter the rendered pad while a hull still reports ~0 separation. World AABBs of the collision meshes and the visual meshes overlap, so this is not a large frame offset.
   • Triangle mesh (physics:approximation = none): PhysX rejects it on these dynamic bodies and logs falling back to convexHull. Allowed dynamic approximations in that log: convex hull, convex decomposition, box, sphere, or SDF.
   • SDF (PhysxSDFMeshCollisionAPI, sdfResolution = 256): the runtime loads approx=sdf and does not fall back to convex hull, but there are zero link3*–bread contacts. The rendered tip keeps moving in until the palm / knuckle hulls hit.

Question

What collision approximation is intended for the Galbot G1 gripper distal links?

convexHull of the visual mesh is what already produces contacts on link1 / link2. SDF on the same instance proxies loads but never reports contact with the bread. Is the remaining visual clip the expected result of convex-hull cooking, or are the distal fingertip colliders not supposed to be generated this way?

Happy to attach the left-wrist frames and the contact log (link name, other body, separation).