Datasets:

---

SomaStacksOrg

/

roboverse_data_wuji

like 0

Follow

SomaStacks Inc 9

Teleop shows fingers sinking INTO objects at sharp edges/corners, on both hands, multiple
angles. Two defects in the Wuji hand fragment, present identically in tianji_wuji.xml and
openarm_wuji.xml:

1. ALL twelve link2 collision geoms per robot (5 proximal-middle capsules + the thumb
   softbody pad, x2 hands) were authored contype="2" conaffinity="2". Objects are 1/1, and
   (2&1)|(2&1)=0 -- the finger mid-segments and thumb pad collided with NOTHING in the
   world. Pressing with those segments passed straight through geometry. (The mask's
   apparent intent was to keep link2 off the palm; the object pass-through came with it.)

2. The link2/link3 collision proxies were capsules inset 3.6-5.6 mm from the visual
   meshes (30% of link2 visual vertices outside the capsule), so even the colliding
   segments contacted late and slid off sharp features.

Fix, per hand and finger:

• link2/link3 capsule -> convex hull of the link's own visual mesh (same mesh asset, body
  frame, rgba kept), contype/conaffinity 1/1
• thumb softbody pad mesh: mask 2/2 -> 1/1
• openarm only: 10 explicit pairs palm_link<->finger*_link2 appended to the
  existing block -- on this hand the palm hull wraps the knuckles (4.6-12.6 mm
  rest-pose overlap), which is what the old mask was hiding; tianji's palm is clear
  (0 overlap measured at rest and at 90% curl) and needs no excludes.

Every finger body has an explicit , so masses/inertias are unchanged; nq/nv
unchanged (recorded demo state buffers still fit).

Validation (MuJoCo 3.8.0, raw load AND composed metasim scene):

• rest pose (teleop hard-zero) and 90%-curled pose: 0 penetrating self-pairs on both robots
• 1000-step zero-ctrl stability: |qvel|max 0.02-0.03, all finite
• visual-vs-collision inset on link2/link3: 3.6-5.6 mm -> 0.00 mm
• step cost (bare robot): tianji 0.087 -> 0.125 ms, openarm ~0.11 ms
• grasp regression re-simulating recorded PD torques across 6 tasks x 2 demos: worst
  hand<->object penetration improves in 12/12 cases (e.g. grasp_apple -10.4 -> -3.5 mm),
  no object leaves the table
• composed-scene deep-contact set is a strict subset of the original's (the remaining
  pairs are the openarm layout's pre-existing rest-hands-on-desk overlap)

Caveat: demos recorded before this change embedded the non-colliding mid-segments inside
object volumes (measured up to 14 mm in an open_cabinet_12 take); state-replay verification
is unaffected, but open-loop torque re-simulation of old takes starts interpenetrated and
depenetrates in the first frames. Live teleop is the intended acceptance test.

tianji_wuji.xml base includes the fused-hand visual graft already on main; the diff is
collision-only.