Grasping by the concavities in a silhouette

A parallel jaw wants two opposing surfaces. For a handle or a rod, the pair is implied by the long axis and grasping-linear-feature finds it. For a part with no long axis — a pair of pliers, a spur gear, a wire spool, a bulb socket — the opposing surfaces are the waists: the places where the outline turns inward. This skill finds them directly.

What the script does

1. Projects the segmented cloud to the (x, y) plane the part is seen from.
2. Builds the alpha shape of that projection — a Delaunay triangulation filtered to triangles whose circumradius is under alpha_radius, whose boundary is the part's real outline rather than its hull.
3. Walks that boundary and keeps the reflex vertices, where it turns inward. Consecutive reflex vertices are one run; each run's centroid is one candidate pad site.
4. Pairs the candidates, keeps the pairs whose secant fits the gripper's span, and builds a straight-down pose centred on each pair's midpoint with the jaws closing along the secant.
5. Ranks by confidence and returns the best, with the rest in candidates so a caller that fails one can try the next without recomputing.

Where it fits

It is a proposer that also poses: it returns grasp_pose and pregrasp_pose directly rather than handing a list to a selector. Put verifying-grasps after it — this skill exits posed with the hand open and does not close it — and route failed to whatever recovery the graph has, because a part whose silhouette has no fitting pair of waists is one this skill genuinely cannot take.

A sensible arrangement is to try grasping-linear-feature first and fall through to this on failed: most parts have an axis, and the ones that do not are exactly the ones this skill is for.