Proposing Short Axis Grasps

proposing-short-axis-grasps

One strategy for the propose stage of a pick: a top-down grasp across the narrow dimension of an elongated body.

Why this is its own skill

A pick has stages — propose, select, approach, close — and each has more than one strategy. Treating "grasping" as one indivisible thing is what leaves a policy with a single approach and nowhere to go when it does not fit: it can only jitter the numbers of a strategy that structurally cannot work. Proposing is separated from selecting so that a refusal here ("the short axis is wider than the hand") is a routable exit to a different rung, not a failed episode.

When to use

• The target is longer than it is wide: a handle, spanner, screwdriver, bar.
• Top-down access is available.
• Any privilege tier. The box may come from sim.get_object_obb or from geometry.compute_obb over a segmented cloud — this skill cannot tell which, which is what makes it portable to the vision body.

When NOT to use

• The top is blocked (a shelf, a lid) — use proposing-side-grasps.
• The body is roughly cubic; there is no meaningful short axis to close across.
• The graspable part is a concave interior (the inside of a ring, a pivot between two plates). A box has no opinion about concavity, and neither has this skill.

Recommended subgraph state flow

describe → get_obb → propose → frame → select → END

1. describe — tool: robot.describe_gripper (for the jaw span).
2. get_obb — tool: sim.get_object_obb, inputs: {object_name: …}. At camera_only/none, substitute geometry.filter_and_compute_obb.
3. propose — script: scripts/<sg>/propose_short_axis.py, inputs target_obb = Ref("get_obb.obb"), gripper = Ref("describe"). Router on route: proposed → frame, declined → the subgraph's declined exit.
4. frame — tool: robot.grasp_frame, inputs: {close_heading_deg: Ref("propose.close_heading_deg")}.
5. select — hand propose.positions and frame.rotation to selecting-reachable-grasp, which walks them.

Required end states