fix(assets): head-heavy mixed-material hammer + recalibrated nail

assets/local_hammer_nail/board_nail.urdf

@@ -6,7 +6,17 @@
      (nail_drive). Used for RENDERING only (physics stays MuJoCo in the hybrid
      path); the joint must exist so metasim can mirror the nail's slide into the
      IsaacSim render each frame. Geometry matches the MJCF (board 0.14x0.14x0.03,
-     nail shaft r0.003 len0.06 + head r0.009 len0.005, joint origin at board top). -->
+     nail shaft r0.003 len0.06 + head r0.009 len0.005, joint origin at board top).
+
+     The <dynamics> values below are INERT: nothing integrates this joint (metasim
+     mirrors the nail's MuJoCo slide into the render each frame), so they are stale
+     copies of the MJCF's pre-2026-07-29 numbers and are deliberately left alone —
+     the generated usd/board_nail.usd was converted from them, and changing one
+     without regenerating the other just splits the pair. board_nail.xml is the
+     authoritative nail model (now frictionloss=60 plus a stiffened friction
+     constraint); URDF has no solreffriction/solimpfriction equivalent at all, so a
+     real IsaacSim-*physics* run of this articulation would still show the
+     soft-constraint creep the MJCF fixes. Physics stays MuJoCo. -->
 <robot name="board_nail">
   <link name="board">
     <visual>

assets/local_hammer_nail/board_nail.xml

@@ -2,23 +2,41 @@
   <!-- Board + nail for the hammer_insert_nail task. The nail is NOT a freejoint
        body: it hangs off the fixed board by a single vertical SLIDE joint with
        dry friction (frictionloss), the standard rigid-body "drive a nail into
-       wood" trick (Adroit hammer task). Static friction (15 N) holds the nail
-       against gravity; a hammer strike whose impulse exceeds it slides the nail
-       down its axis a little — visually "driven in". Because the joint only
-       translates along z, the nail can NOT be knocked flat, so this is a genuine
-       vertical-insertion task (vs hammer_drive_peg's knock-flat).
+       wood" trick (Adroit hammer task). Static friction (60 N) holds the nail
+       against gravity and against being leaned on; a hammer strike whose impulse
+       exceeds it slides the nail down its axis a little — visually "driven in".
+       Because the joint only translates along z, the nail can NOT be knocked flat,
+       so this is a genuine vertical-insertion task (vs hammer_drive_peg's knock-flat).
 
        Modeling notes:
        - The nail SHAFT has collision off (contype=0/conaffinity=0): its going
          "into the wood" is faked by the slide joint, so it must not fight real
          contact with the board. Only the nail HEAD keeps collision — that is the
          surface the hammer strikes.
-       - damping on the joint spreads the drive over several strikes (one blow
-         does not bottom it out), matching a real hammering rhythm.
+       - damping on the joint spreads the drive over several strikes (one blow does
+         not bottom it out), matching a real hammering rhythm.
+       - solreffriction + solimpfriction stiffen the FRICTION constraint, and both
+         are load-bearing, not cosmetic. MuJoCo's frictionloss is a SOFT constraint:
+         a *sustained* load deforms it by an amount set by the constraint impedance
+         and NOT by the frictionloss magnitude, so the nail creeps down with no
+         strike at all. solreffriction's timeconst is already at its stability floor
+         (2*dt = 0.004 at the family's dt=0.002) and alone left ~5 mm of gravity sag
+         and ~30 mm under a 0.665 kg hammer simply laid on the nail head — i.e. past
+         the checker's 25 mm gate with no blow struck (measured 2026-07-29, after the
+         YCB hammer's head-heavy mass fix put the tool's real weight on the nail).
+         solimpfriction hardens it the rest of the way: gravity sag 0.05 mm, a laid-on
+         hammer 1.5 mm, a hard 30 N press 6.6 mm. Do NOT loosen either.
+       - frictionloss (60 N) is what a strike must beat to slide the nail, so it sets
+         the drive per blow. Calibrated against the corrected head-heavy hammer
+         (effective striking mass 599 g of 665 g): ~1 mm per 1.0 m/s tap (never
+         latches), 6.5 mm per 1.5 m/s swing (latches on blow 4), 12.8 mm per 2.0 m/s
+         blow (latches on blow 3). The pre-fix handle-heavy hammer only delivered
+         331 g into the strike, which is why 15 N used to look right.
        - Loaded as ArticulationObjCfg(fix_base_link=True); the root body carries
          no pos/quat (set via the cfg default_position), as in desk_lamp.xml.
        - Success is read from the nail_drive joint position (insertion depth):
-         qpos 0 = flush with the start, range floor -0.04 = fully driven. -->
+         qpos 0 = flush with the start, range floor -0.04 = fully driven; the hammer
+         head bottoms out on the board at ~29.6 mm (the nail is then driven flush). -->
   <compiler angle="radian"/>
   <worldbody>
     <body name="board">
@@ -26,7 +44,8 @@
             rgba="0.55 0.38 0.22 1" friction="0.9 0.05 0.001" density="500"/>
       <body name="nail" pos="0 0 0.03">
         <joint name="nail_drive" type="slide" axis="0 0 1" range="-0.04 0"
-               frictionloss="15" damping="10" armature="0.05"/>
+               frictionloss="60" damping="10" armature="0.05"
+               solreffriction="0.004 1" solimpfriction="0.999 0.9999 0.0001 0.5 2"/>
         <geom name="nail_shaft" type="cylinder" size="0.003 0.03" pos="0 0 0.0"
               density="7850" contype="0" conaffinity="0" rgba="0.72 0.72 0.75 1"/>
         <geom name="nail_head" type="cylinder" size="0.009 0.0025" pos="0 0 0.032"

assets/ycb/hammer/hammer.xml

@@ -5,7 +5,12 @@
   textured.obj -> trimesh recenter (xy centered on AABB, base at z=0) -> visual mesh
   + CoACD convex-decomposed collision (7 parts, threshold=0.04) so the
   concave shape is collision-true (cavity preserved / tool silhouette kept).
-  Density 1840.5 kg/m^3 over the summed part volume 0.000361 m^3 -> mass 0.665 kg.
+  Mixed-material mass (hammer is a two-material tool; a single density would put the
+  centre of mass at the volume centroid): 531 g dense (implied 6663 kg/m^3 over the 79.6 cm^3 the scan texture reads as
+  bare metal, vs 7850 for solid steel) + 134 g light (750 kg/m^3 x 179.3 cm^3) -> mass 0.665 kg,
+  CoM [-0.0223, 0.089, 0.0166]. Per-geom density is an EFFECTIVE density over the
+  convex-hull part volume (361.3 cm^3 vs the scan's 258.9 cm^3), so it
+  reads lower than the material it stands for — do not "correct" it back to 7850.
   Recentered AABB (m): size=[0.1822, 0.3327, 0.0329].
   License: YCB CC BY 4.0.
 -->
@@ -28,25 +33,25 @@
       <geom name="hammer_visual" type="mesh" mesh="hammer_mesh" material="hammer_mat"
             group="2" contype="0" conaffinity="0" density="0" />
       <geom name="hammer_collision0" type="mesh" mesh="hammer_col0"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="547.4" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision1" type="mesh" mesh="hammer_col1"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="4717.1" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision2" type="mesh" mesh="hammer_col2"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="2067.7" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision3" type="mesh" mesh="hammer_col3"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="4409.4" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision4" type="mesh" mesh="hammer_col4"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="4717.1" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision5" type="mesh" mesh="hammer_col5"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="4717.1" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
       <geom name="hammer_collision6" type="mesh" mesh="hammer_col6"
-            density="1840.5" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
+            density="4717.1" friction="2.0 0.1 0.01" solref="0.01 1" priority="2"
             group="3" rgba="0 1 0 0" />
     </body>
   </worldbody>