# Verifying Grasps

> Three independent checks on one grasp attempt -- that the hand arrived where it was sent before the jaws close, that the jaw gap after the close is not the mechanical stop, and that the lifted object is visible above the surface and near the hand. Use when a grasp closed on a small or thin object and the graph must know, before carrying or inserting, whether anything is actually in the gripper.

- Skill: `graph-robots/verifying-grasps` (Agent Skill, multi-file: 4 files)
- Install (CLI): `npx skillmds@latest add graph-robots/verifying-grasps`
- Raw SKILL.md: https://api.skillmd.com/api/skills/graph-robots/verifying-grasps/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: Apache-2.0
- Author: graph-robots (https://skillmd.com/u/graph-robots)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/graph-robots/verifying-grasps

---


# verifying-grasps

Three instruments on one grasp attempt, at three different moments. A graph may
use any of them alone; together they separate "the hand never got there" from
"it got there and closed on nothing" from "it closed on something and dropped
it", which a single check cannot.

| when | script | routes |
| --- | --- | --- |
| after the approach, before the close | `verify_reach` | `reached` / `missed` |
| immediately after the close | `verify_grip` | `held` / `empty` / `unknown` |
| after a short lift | `verify_grasp` | `verified` / `not_held` |

`verify_reach` compares the commanded grasp pose with the observed one and
routes `missed` when the distance exceeds the object's own smallest half-extent
(clamped to `floor_m`..`ceiling_m`). The tolerance is the object's size rather
than a constant because the question is whether the object can still be between
the pads, and that is a question about the object. Cheap -- no camera, no
motion -- and it fires before the close, when the graph can still do something
about it.

`verify_grip` reads `robot.get_gripper` and converts through
`robot.describe_gripper`'s `width_fit`, routing `empty` when the gap is within
`empty_margin_m` of the mechanical stop. Also cameraless. It reports
`expected_width_m` alongside, which is context and not a test: a planner that
grasps at a fitted line's centre may legitimately hold a wide part of the body.

`verify_grasp` raises the end effector by `lift_m` along world Z with a
Cartesian move, takes a fresh observation, and looks for `object_description`
(then `marker_description`, when given) first in the camera whose name contains
`wrist_camera_keyword` and then in `overhead_camera_name`. The first
segmentation at or above `score_min` is back-projected; the grasp is
`verified` only when the cloud has at least `min_points` valid depth points,
its median lies at least `min_above_table_m` above
`robot.describe_workspace().surface_z`, and within `max_hand_distance_m` of
the lifted hand. Every failed gate returns `route: not_held` with a `reason`.

`score_min` is deliberately low: a small held object occupies only a few
dozen wrist pixels and true positives score in the hundredths. Confidence
admits a candidate; geometry decides.

## Boundaries

- The lift in `verify_grasp` is the only motion any of the three makes; the
  object is neither released nor moved elsewhere.
- No force or tactile reading is consulted. `verify_grip` uses the jaw gap,
  which every parallel hand reports; torque and slip belong to a connector's
  own grasp check when it has one.
- Only a rig without a wrist camera raises, and only from `verify_grasp`. A
  missing input elsewhere routes the benign way -- `reached`, `unknown` --
  because absent evidence is not evidence of a failed grasp.