feat(maniskill): port PegInsertionSide dense reward — 10/10 native dense rewards bitwise

reward staged (reach peg -> grasp -> pre-insert align -> insert) computed from peg/goal/
box-hole poses; matches native compute_dense_reward to float32 eps (5.96e-8). Wired into
the shipped peg_insertion_side_native task. All 10 tasks with a native dense reward are now
bitwise; all 12 tabletop tasks have bitwise success.

Author: geng-haoran

roboverse_pack/tasks/maniskill/_native/rewards.py

@@ -248,3 +248,21 @@ def pull_cube_tool(
     if success:
         reward += 5.0
     return float(reward)
+
+
+def peg_insertion_side(*, gripper_to_peg, is_grasped, yz1, yz2, hole_dist, success, normalized=False) -> float:
+    """ManiSkill PegInsertionSide-v1 ``compute_dense_reward`` (max 10).
+
+    Scalars are computed by the caller from poses (gripper→peg distance; yz1/yz2 = |yz| of peg-head /
+    peg-center in goal frame; hole_dist = |peg head in box-hole frame|; is_grasped at max_angle=20).
+    """
+    reward = 1.0 - _tanh(4.0 * gripper_to_peg)
+    reward += 1.0 if is_grasped else 0.0
+    pre = 3.0 * (1.0 - _tanh(0.5 * (yz1 + yz2) + 4.5 * max(yz1, yz2)))
+    reward += pre * (1.0 if is_grasped else 0.0)
+    pre_inserted = yz1 < 0.01 and yz2 < 0.01
+    insertion = 5.0 * (1.0 - _tanh(5.0 * hole_dist))
+    reward += insertion * (1.0 if (is_grasped and pre_inserted) else 0.0)
+    if success:
+        reward = 10.0
+    return float(reward / 10.0) if normalized else float(reward)

roboverse_pack/tasks/maniskill/_native/specs.py

@@ -323,6 +323,30 @@ def _pull_cube_tool_success(task):
     return SU.pull_cube_tool(cube_pos=task.obj_pos("cube"), robot_base_pos=task.robot_base_pos())
 
 
+def _peg_insertion_reward(task):
+    import numpy as _np
+    import sapien
+
+    box = task.handler.object_ids["box_with_hole_0"].get_pose()
+    peg = task.handler.object_ids["peg_0"].get_pose()
+    goal = box * sapien.Pose([-0.10695, 0.0087, 0.0038], [1, 0, 0, 0])
+    ph = peg * sapien.Pose([0.10695, 0.0, 0.0], [1, 0, 0, 0])
+    bh = box * sapien.Pose([0.0, 0.0087, 0.0038], [1, 0, 0, 0])
+    tgt = peg * sapien.Pose([-0.06, 0.0, 0.0], [1, 0, 0, 0])
+    gtp = float(_np.linalg.norm(task.tcp_pos() - _np.asarray(tgt.p)))
+    yz1 = float(_np.linalg.norm(_np.asarray((goal.inv() * ph).p)[1:]))
+    yz2 = float(_np.linalg.norm(_np.asarray((goal.inv() * peg).p)[1:]))
+    hd = float(_np.linalg.norm(_np.asarray((bh.inv() * ph).p)))
+    return RW.peg_insertion_side(
+        gripper_to_peg=gtp,
+        is_grasped=task.is_grasped("peg_0", max_angle=20),
+        yz1=yz1,
+        yz2=yz2,
+        hole_dist=hd,
+        success=_peg_insertion_success(task),
+    )
+
+
 def _peg_insertion_success(task):
     """ManiSkill PegInsertionSide ``has_peg_inserted``: peg head inside the box hole.
 
@@ -545,6 +569,7 @@ def _stack_pyramid_success(task):
         "success": _moved("peg_0", (-0.0007, -0.0695)),
         "orientations": {"box_with_hole_0": [0.6694, 0.0, 0.0, 0.7429], "peg_0": [0.8344, 0.0, 0.0, 0.5511]},
         "success_full": _peg_insertion_success,
+        "reward": _peg_insertion_reward,
     },
     "plug_charger": {
         "gym_id": "PlugCharger-v1",