feat(juice_shot): v3 — gate at visible disengagement, longer travel, easier twist

assets/local_juice_shot/juice_shot.urdf

@@ -12,15 +12,16 @@
      fix_base_link=False, matching the MuJoCo freejoint.
 
      SCREW COUPLING CAVEAT (same as bottle_cap.urdf): URDF cannot express the
-     MJCF joint equality (cap_slide = 0.000477465 * cap_twist, 3 mm/turn), so
+     MJCF joint equality (cap_slide = 0.00128916 * cap_twist, 8.1 mm/turn), so
      this twin chains the two joints WITHOUT the coupling; wiring it on the
      Isaac side is part of the documented one-time USD conversion follow-up.
      MuJoCo — the physics backend of record — is unaffected.
 
      Real-world reference (sim2real): Trader Joe's Organic Mighty Turmeric
      juice shot, 2 fl oz / 59 mL — ~85 mm tall, 38 mm body, 29 mm white ribbed
-     cap, 33 mm support flange, thread 3.0 mm/turn (28 mm PCO-1881 reference
-     2.7 mm), ~1.7 turns to open, ~75 g filled. -->
+     cap, 33 mm support flange, 3-START thread with a 2.7 mm ridge-to-ridge
+     pitch => LEAD 8.1 mm of rise per turn (28 mm PCO-1881 finish), ~0.6 turn
+     to open, ~75 g filled. -->
 <robot name="juice_shot">
   <link name="bottle">
     <inertial>
@@ -81,7 +82,7 @@
     <child link="cap_screw"/>
     <origin xyz="0 0 0.069"/>
     <axis xyz="0 0 1"/>
-    <limit lower="0.0" upper="16.7552" effort="2" velocity="20.0"/>
+    <limit lower="0.0" upper="62.832" effort="2" velocity="20.0"/>
     <dynamics damping="0.005" friction="0.02"/>
   </joint>
   <joint name="cap_slide" type="prismatic">
@@ -89,7 +90,7 @@
     <child link="cap"/>
     <origin xyz="0 0 0"/>
     <axis xyz="0 0 1"/>
-    <limit lower="0.0" upper="0.008" effort="50" velocity="0.5"/>
+    <limit lower="0.0" upper="0.040" effort="50" velocity="0.5"/>
     <dynamics damping="0.2" friction="0.0"/>
   </joint>
 </robot>

assets/local_juice_shot/juice_shot.xml

@@ -1,5 +1,5 @@
 <mujoco model="juice_shot">
-  <!-- 59 mL juice-shot bottle + FINE-PITCH screw cap for the open_juice_shot
+  <!-- 59 mL juice-shot bottle + 3-START screw cap for the open_juice_shot
        task — a sim2real twin of a real retail bottle (Trader Joe's Organic
        Mighty Turmeric juice shot, NET 2 fl oz / 59 mL), dimensioned from the
        maintainer's reference photos + standard PET shot-bottle specs:
@@ -7,23 +7,70 @@
          total height (closed)  ~85 mm     body OD          ~38 mm
          cap OD                 ~29 mm     cap height       ~15.5 mm
          support flange OD      ~33 mm     tamper band      stays on the neck
-         thread pitch           3.0 mm/turn, single-start FINE
-                                (reference: 28 mm PCO-1881 beverage finish =
-                                2.7 mm/turn), ~1.7 turns to disengage
+         thread                 3-START, 2.7 mm ridge-to-ridge pitch =>
+                                LEAD 8.1 mm of cap rise per turn (28 mm
+                                PCO-1881 beverage finish), ~0.6 turn to
+                                disengage — the half-turn open of a real bottle
          filled mass            ~75 g (59 g juice + ~11 g PET + ~5 g cap)
 
        The cap is the same screw mechanism as local_bottle_cap v2 — SLIDE
        (cap_slide, listed FIRST for the parrot_sim overlay's first-joint
-       contract) + coaxial HINGE (cap_twist), tied by the joint equality below:
-       slide = PITCH * twist, PITCH = 0.003 / 2π = 0.000477465 m/rad. The only
-       way the cap rises is by TURNING counter-clockwise: ~1.7 turns to the
-       task's 5 mm gate, ~2.7 turns to the 8 mm stop. Dry thread friction
-       (0.02 N·m) means a straight pull must back-drive the FINE screw at
-       ≈ frictionloss / PITCH ≈ 42 N — the finer pitch gives the thread MORE
-       mechanical advantage than the big bottle's 18 mm/turn, so this tiny cap
-       is even harder to yank off, exactly as with real beverage closures
-       (measured: 20 N pulls deflect ~0.45 mm elastically; ~42 N+ slips the
-       thread and the cap backs off ROTATING, like a stripped real cap).
+       contract) + coaxial HINGE (cap_twist), tied by the named joint equality
+       "cap_screw" below: slide = LEAD * twist, LEAD = 0.0081 / 2π =
+       0.00128916 m/rad — WITH a v3 twist-then-PULL-OFF release on top.
+
+       v3 release mechanism (2026-08-10, operator feedback, second iteration —
+       "I want the cap completely out of the bottle"):
+         * ENGAGED (slide < 8 mm): the screw equality binds rise to twist —
+           the only way up is turning ccw, a straight pull must back-drive the
+           thread at ~19 N. ~1 full turn covers the engaged band — exactly the
+           travel the first 19 recorded takes delivered (5.4-8.1 mm, two
+           parked on the old 8 mm stop).
+         * RELEASED (slide >= 8 mm): the TASK deactivates "cap_screw"
+           (data.eq_active, a stateless per-step rule wrapped around
+           handler.simulate, so teleop / collect / eval all share it — the
+           MJCF alone cannot express a conditional equality). The cap slides
+           freely: the operator PULLS it straight off, like a real bottle
+           after its last thread lets go. Dropping it lets it sink back and
+           re-seat (armature 1.0 slows the fall; re-engagement is seamless
+           when the cap was not spun while free).
+         * SUCCESS (task gate, min_open_m = 0.015): cap risen 15 mm — its
+           skirt bottom (0.069 + rise) is then 2 mm ABOVE the entire neck
+           finish (lip at 0.082), i.e. the cap is visibly clear of the whole
+           bottle, held in the fingers. Slide range tops at 40 mm so the
+           lift-off has real headroom; twist is UNLIMITED (a freed spinning
+           cap must not hit a phantom stop).
+
+       History of the gate: v2 latched at 5 mm — kinematically where a 3-start
+       closure disengages, but the skirt still covered 5 mm of visible thread
+       (SUCCESS looked premature) and the 8 mm stop truncated the unscrew.
+       The first v3 draft moved the gate to 10 mm on a longer screw; operator
+       feedback pushed further: the point is not "more screwing", it is the
+       cap COMING OFF — hence the release mechanism instead of more turns.
+
+       Thread friction is 0.025 N·m (operator-chosen; breakaway headroom vs
+       the measured >=0.047 N·m: 1.88x). The cap knurl grip is 4.0 (ribbed
+       interlock; grip friction is the sanctioned "easier to twist" lever —
+       it multiplies deliverable torque at the 14.5 mm radius without
+       weakening the thread's pull resistance).
+
+       LEAD, NOT PITCH — this is the modelling point of the asset. PCO-1881 is
+       a 3-START thread: ridges are 2.7 mm apart along the axis, but one turn
+       advances the cap past all three of them, so it rises 8.1 mm per turn and
+       the closure disengages in about half a turn. That is why a real bottle
+       opens with ONE wrist rotation. Treating the 2.7 mm pitch as the lead
+       (v1 of this asset, single-start 3 mm/turn) models a closure 3x finer
+       than the bottle it is a twin of, and turns "open the bottle" into ~1.7
+       turns — 4 regrasps for a teleoperator, which is why not one of the first
+       22 recorded takes reached the gate (best 475° = 84% of it).
+
+       Yank guard (engaged band): a straight pull must back-drive the screw
+       at ≈ frictionloss / LEAD ≈ 19 N sustained through the whole 8 mm
+       engaged band (~0.16 J) before the release frees the cap — past a
+       fingertip pinch, and the settle gate still demands a controlled hold
+       at the 15 mm gate (a ballistic pop is rejected). Real caps strip too;
+       the guard's job is pricing the cheat above honest twisting, not
+       making it impossible.
        The BOTTLE BASE IS FREE (cfg fix_base_link=False): at ~75 g the bottle
        spins with the cap unless the other hand pins it — bimanual by
        mechanics, and light enough that a table-friction-only "no-hands" cheat
@@ -35,15 +82,25 @@
          is SKIPPED when the parent welds to the world, i.e. exactly in the
          fixed-base standalone loads the asset regression tests use;
        - solreffriction stiffens the friction constraint (no creep: cap weight
-         back-drives at ~4e-5 N·m, 500x below the 0.02 N·m thread friction);
-       - collision = primitive cylinders (group 3, hidden) with the calibrated
-         grasp friction/priority; visuals = group-1 meshes from
-         gen_juice_shot_meshes.py (deterministic; the visible thread helix is
-         generated at the SAME 3 mm/turn pitch as the equality);
+         back-drives at ~1.2e-4 N·m, 208x below the 0.025 N·m thread friction —
+         measured settle -0.003 mm over 4 s);
+       - collision = primitive cylinders (group 3, hidden): the bottle body
+         carries the pack's authored grasp friction 3.0 0.15 0.01 + priority 2
+         (the same tuple the box_task props use); the CAP knurl carries 4.0
+         since v3 (ribbed interlock). The ribbed cap and the pinned body are
+         BOTH grasp targets here, and grip friction is the lever that gives the operator
+         twisting torque WITHOUT weakening the thread's yank resistance: a
+         higher mu multiplies the tangential force a pinch can deliver at the
+         14.5 mm cap radius, while the straight-pull threshold stays a property
+         of the thread (friction / lead). Visuals = group-1 meshes from
+         gen_juice_shot_meshes.py (deterministic; the visible helix is
+         generated as the SAME 3 starts at the SAME 8.1 mm lead as the
+         equality, so what is drawn is the screw that is simulated);
        - the juice visual cylinder carries the liquid mass (density 1000);
        - success is the cap's base-relative displacement (articulated_open,
-         min_open_m = 0.005): qpos 0 = closed, slide range top 0.008;
-         cap origin (0 0 0.069) == the checker's closed_rel_pos_m;
+         min_open_m = 0.015): qpos 0 = closed, release at 0.008 (task rule),
+         slide range top 0.040; cap origin (0 0 0.069) == the checker's
+         closed_rel_pos_m;
        - loaded as ArticulationObjCfg(fix_base_link=False); the root body
          carries no pos/quat (set via the cfg default_position). -->
   <compiler angle="radian" meshdir="meshes"/>
@@ -64,13 +121,13 @@
     <body name="bottle">
       <!-- collision (hidden group 3) -->
       <geom name="shot_body_col" type="cylinder" size="0.019 0.026" pos="0 0 0.026" group="3"
-            density="220" friction="2.0 0.1 0.005" priority="2" rgba="0.2 0.2 0.2 0"/>
+            density="220" friction="3.0 0.15 0.01" priority="2" rgba="0.2 0.2 0.2 0"/>
       <geom name="shot_neck_col" type="cylinder" size="0.0125 0.0075" pos="0 0 0.0595" group="3"
-            density="800" friction="2.0 0.1 0.005" priority="2" rgba="0.2 0.2 0.2 0"/>
+            density="800" friction="3.0 0.15 0.01" priority="2" rgba="0.2 0.2 0.2 0"/>
       <geom name="shot_flange_col" type="cylinder" size="0.0165 0.001" pos="0 0 0.0665" group="3"
-            density="800" friction="2.0 0.1 0.005" priority="2" rgba="0.2 0.2 0.2 0"/>
+            density="800" friction="3.0 0.15 0.01" priority="2" rgba="0.2 0.2 0.2 0"/>
       <geom name="shot_topneck_col" type="cylinder" size="0.0122 0.0075" pos="0 0 0.0745" group="3"
-            density="600" friction="2.0 0.1 0.005" priority="2" rgba="0.2 0.2 0.2 0"/>
+            density="600" friction="3.0 0.15 0.01" priority="2" rgba="0.2 0.2 0.2 0"/>
       <!-- visual -->
       <geom name="shot_juice_vis" type="cylinder" size="0.0163 0.0225" pos="0 0 0.0275"
             material="mat_juice" contype="0" conaffinity="0" group="1" density="1000"/>
@@ -88,19 +145,27 @@
             contype="0" conaffinity="0" group="1" density="100"/>
       <body name="cap" pos="0 0 0.069">
         <!-- slide FIRST (overlay contract — see header) -->
-        <!-- The high slide armature models the fine screw's TRANSLATIONAL
-             REFLECTED INERTIA (I_twist / pitch² ≈ 440 kg-equivalent for this
-             thread — 1.0 is conservative): a yank meets thread mechanical
-             advantage inertially, and gravity settle tightens 0.10 → 0.003 mm.
-             It does NOT fix the equality's elastic leak under load — that
-             lever is the equality solimp (measured; see the equality comment).
-             Twist-side cost is negligible (armature · pitch² ≈ 2e-7 kg·m²). -->
-        <joint name="cap_slide" type="slide" axis="0 0 1" range="0 0.008"
+        <!-- The high slide armature models the screw's TRANSLATIONAL REFLECTED
+             INERTIA (I_twist / lead² ≈ 60 kg-equivalent at this 3-start lead,
+             was 440 at the single-start one — 1.0 stays conservative either
+             way): a yank meets thread mechanical advantage inertially, and
+             gravity settle tightens 0.10 → 0.003 mm. It does NOT fix the
+             equality's elastic leak under load — that lever is the equality
+             solimp (measured; see the equality comment). Twist-side cost is
+             negligible (armature · lead² ≈ 1.7e-6 kg·m²). -->
+        <joint name="cap_slide" type="slide" axis="0 0 1" range="0 0.040"
                damping="0.2" armature="1.0"/>
-        <joint name="cap_twist" type="hinge" axis="0 0 1" range="0 16.7552"
-               frictionloss="0.02" damping="0.005" armature="0.0001" solreffriction="0.004 1"/>
+        <!-- Twist range (v3): the FLOOR at 0 is what blocks the tighten
+             direction (the slide floor alone loses to the much stiffer screw
+             equality), and the 200 rad ceiling (~32 turns) is unlimited for
+             any real take — once the task's release rule disengages the
+             equality (slide >= 8 mm) a spinning freed cap must not hit a
+             phantom stop. frictionloss sets the straight-pull threshold
+             while engaged, ≈ 19 N — see the header. -->
+        <joint name="cap_twist" type="hinge" axis="0 0 1" range="0 200"
+               frictionloss="0.025" damping="0.005" armature="0.0001" solreffriction="0.004 1"/>
         <geom name="shot_cap_col" type="cylinder" size="0.0145 0.00775" pos="0 0 0.00775" group="3"
-              density="800" friction="2.0 0.1 0.005" priority="2" rgba="0.2 0.2 0.2 0"/>
+              density="800" friction="4.0 0.15 0.01" priority="2" rgba="0.2 0.2 0.2 0"/>
         <geom name="shot_cap_vis" type="mesh" mesh="shot_cap_mesh" material="mat_cap_white"
               contype="0" conaffinity="0" group="1" density="100"/>
       </body>
@@ -110,14 +175,14 @@
     <exclude body1="bottle" body2="cap"/>
   </contact>
   <equality>
-    <!-- The fine screw: cap_slide = 0.000477465 * cap_twist (3 mm per 2π turn). -->
+    <!-- The screw: cap_slide = 0.00128916 * cap_twist (8.1 mm per 2π turn). -->
     <!-- solimp is deliberately extreme: with a 10 g cap and the solref
          timeconst floored at 2*dt, the equality's REGULARIZATION (efc_R,
          set by 1-impedance) is what a pull deflects against — measured, a
          20 N pull "opened" 3.2 mm at the bottle_cap-style 0.99 impedance
          with the thread friction NOT slipping. 0.999/0.9999 brings that
          elastic leak to 0.45 mm while twist-opening is unaffected. -->
-    <joint joint1="cap_slide" joint2="cap_twist" polycoef="0 0.000477465 0 0 0"
+    <joint name="cap_screw" joint1="cap_slide" joint2="cap_twist" polycoef="0 0.00128916 0 0 0"
            solref="0.004 1" solimp="0.999 0.9999 0.0005"/>
   </equality>
 </mujoco>