grasping-short-axis

Name: grasping-short-axis
Author: graph-robots

26stars4forks

How to install grasping-short-axis?

graph-robots/open-robot-skills/grasping-short-axis

$npx -y skills add graph-robots/open-robot-skills --skill grasping-short-axis

Installs into the current project.

Summary

--- name: grasping-short-axis description: Deterministic short-axis-aligned grasp with CuRobo. The grasp pose is computed geometrically (NOT sampled/scored) — the gripper approaches straight down with its finger-opening axis locked to the OBB's SHORTER horizontal axis, so t

Files · 1

View on GitHub

SKILL.md

1	---
2	name: grasping-short-axis
3	description: Deterministic short-axis-aligned grasp with CuRobo. The grasp
4	pose is computed geometrically (NOT sampled/scored) — the gripper
5	approaches straight down with its finger-opening axis locked to the OBB's
6	SHORTER horizontal axis, so the jaws close ACROSS the narrow dimension of
7	an elongated target. Use for pan/pot handles, bottles, tools, utensils —
8	anywhere the OBB centroid is graspable but orientation is the thing that
9	matters. An optional offset_from_base node slides the grasp outward along
10	the handle, clear of the heavier object body, when a second OBB of that
11	body is wired in. A scored/sampled pose often holds at close but shears
12	out under the lift; the short-axis pose grips the bar squarely so it
13	survives lift + transport.
14	compatibility: requires gap>=0.1
15	metadata: {category: grasping, tags: [grasping, manipulation, collision-aware, curobo, short-axis, handle]}
16	gap:
17	allowed_tools:
18	- geometry.build_world_config
19	- robot.get_observation
20	- curobo.plan_to_grasp_poses
21	- curobo.plan_grasp_motion
22	- curobo.plan_directed_linear
23	- robot.execute_trajectory
24	- robot.move_to_joints
25	- robot.get_ee_pose
26	- robot.go_to_pose
27	- robot.go_to_pose_cartesian
28	- robot.open_gripper
29	- robot.close_gripper
30	exit_conditions:
31	grasped: Object held in the gripper after `close`.
32	failed: Any failure during the grasp attempt — collision-aware planning failure or trajectory execution error (a raise propagating to `on_error`). Coordinator routes to abort.
33	required_inputs:
34	# MANDATORY — the coordinator MUST declare both in `grasp_sg.inputs`
35	# and wire them from the upstream handle-perception subgraph.
36	# OPTIONAL extra input (NOT listed here so graph-validation does not
37	# force it): `base_obb: OrientedBoundingBox`. See the `offset_from_base`
38	# hard_rule — declare + wire it ONLY when you also authored a second
39	# perception subgraph for the heavier object body (pan bowl / pot
40	# belly). Omitting it is valid; the grasp degrades to a plain
41	# short-axis grasp (still correct, just not slid off the body).
42	target_obb: OrientedBoundingBox
43	target_mask: Mask
44	produces_outputs:
45	ee_pose_at_grasp: Se3Pose
46	grasp_pose: Se3Pose
47	hard_rules:
48	- >
49	ALWAYS begin with `robot.open_gripper` (settle_steps=40). Skipping
50	this is the #1 silent grasp-failure mode when the gripper starts
51	closed from a previous task step.
52	- >
53	Use the canonical `compute_grasp` (`scripts/<sg>/short_axis_grasp_pose.py`)
54	— it returns ONE deterministic Se3Pose aligned to the OBB's shorter
55	horizontal axis. Do NOT swap in `geometry.top_down_grasp_candidates`
56	or any sampled/scored generator: the whole point of this skill is
57	that the orientation is geometry-locked, not discriminator-ranked.
58	`plan_grasp` auto-wraps the single pose into a one-element goalset.
59	- >
60	`offset_from_base` (`scripts/<sg>/offset_grasp_from_base.py`) is the
61	robustness lever for a handle attached to a heavier body (frying pan,
62	pot). It slides the grasp outward along the handle, clear of the
63	body, so the jaws grip the bar — not the rim — and the grip survives
64	the lift. To enable it: (1) author a SECOND perception subgraph that
65	segments the object BODY (e.g. perceiving-object-parts /
66	perceiving-objects with object/subpart name like "the round base of
67	the frying pan"),
68	(2) add `base_obb: OrientedBoundingBox` to THIS subgraph's `inputs`
69	and wire it from that body perception, (3) bind
70	`base_obb = Ref("in.base_obb")` on the `offset_from_base` node.
71	If you do NOT author a body perception, KEEP the `offset_from_base`
72	state but bind only `handle_obb` + `grasp_pose` (omit `base_obb`);
73	the script no-ops and returns the grasp unchanged. NEVER write
74	`Ref("in.base_obb")` unless `base_obb` is a declared subgraph input —
75	that is a graph-validation error.
76	- >
77	When building the collision world, ALWAYS pass
78	`target_mask = Ref("in.target_mask")` AND
79	`target_obb = Ref("in.target_obb")` AND a `target_name`. The mask is
80	the authoritative target silhouette (pixel-accurate); the OBB is a
81	safety fallback. The same `target_name` MUST be passed to `plan`
82	(`plan_grasp.py`) so the planner excludes that mesh from collision.
83	- >
84	ALWAYS include the `approach` state between `offset_from_base` and
85	`observe`. `plan_grasp` disables CuRobo's `use_grasp_approach` flag on
86	the assumption the gripper is already oriented over the target —
87	skipping `approach` makes the descent start from a wrong wrist
88	rotation and slip. `approach_above.py` lifts to a safe height,
89	translates above the target, then rotates in place to the grasp
90	orientation — all three legs route through
91	`robot.go_to_pose_cartesian` (gap.connector.ik.CuRoboBackend), which
92	uses cuRobo's

People who liked this also viewed

6 picks

Type

grasping-short-axis

26stars4forks

View on GitHub

How to install grasping-short-axis?

graph-robots/open-robot-skills/grasping-short-axis

$npx -y skills add graph-robots/open-robot-skills --skill grasping-short-axis

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	---
2	name: grasping-short-axis
3	description: Deterministic short-axis-aligned grasp with CuRobo. The grasp
4	pose is computed geometrically (NOT sampled/scored) — the gripper
5	approaches straight down with its finger-opening axis locked to the OBB's
6	SHORTER horizontal axis, so the jaws close ACROSS the narrow dimension of
7	an elongated target. Use for pan/pot handles, bottles, tools, utensils —
8	anywhere the OBB centroid is graspable but orientation is the thing that
9	matters. An optional offset_from_base node slides the grasp outward along
10	the handle, clear of the heavier object body, when a second OBB of that
11	body is wired in. A scored/sampled pose often holds at close but shears
12	out under the lift; the short-axis pose grips the bar squarely so it
13	survives lift + transport.
14	compatibility: requires gap>=0.1
15	metadata: {category: grasping, tags: [grasping, manipulation, collision-aware, curobo, short-axis, handle]}
16	gap:
17	allowed_tools:
18	- geometry.build_world_config
19	- robot.get_observation
20	- curobo.plan_to_grasp_poses
21	- curobo.plan_grasp_motion
22	- curobo.plan_directed_linear
23	- robot.execute_trajectory
24	- robot.move_to_joints
25	- robot.get_ee_pose
26	- robot.go_to_pose
27	- robot.go_to_pose_cartesian
28	- robot.open_gripper
29	- robot.close_gripper
30	exit_conditions:
31	grasped: Object held in the gripper after `close`.
32	failed: Any failure during the grasp attempt — collision-aware planning failure or trajectory execution error (a raise propagating to `on_error`). Coordinator routes to abort.
33	required_inputs:
34	# MANDATORY — the coordinator MUST declare both in `grasp_sg.inputs`
35	# and wire them from the upstream handle-perception subgraph.
36	# OPTIONAL extra input (NOT listed here so graph-validation does not
37	# force it): `base_obb: OrientedBoundingBox`. See the `offset_from_base`
38	# hard_rule — declare + wire it ONLY when you also authored a second
39	# perception subgraph for the heavier object body (pan bowl / pot
40	# belly). Omitting it is valid; the grasp degrades to a plain
41	# short-axis grasp (still correct, just not slid off the body).
42	target_obb: OrientedBoundingBox
43	target_mask: Mask
44	produces_outputs:
45	ee_pose_at_grasp: Se3Pose
46	grasp_pose: Se3Pose
47	hard_rules:
48	- >
49	ALWAYS begin with `robot.open_gripper` (settle_steps=40). Skipping
50	this is the #1 silent grasp-failure mode when the gripper starts
51	closed from a previous task step.
52	- >
53	Use the canonical `compute_grasp` (`scripts/<sg>/short_axis_grasp_pose.py`)
54	— it returns ONE deterministic Se3Pose aligned to the OBB's shorter
55	horizontal axis. Do NOT swap in `geometry.top_down_grasp_candidates`
56	or any sampled/scored generator: the whole point of this skill is
57	that the orientation is geometry-locked, not discriminator-ranked.
58	`plan_grasp` auto-wraps the single pose into a one-element goalset.
59	- >
60	`offset_from_base` (`scripts/<sg>/offset_grasp_from_base.py`) is the
61	robustness lever for a handle attached to a heavier body (frying pan,
62	pot). It slides the grasp outward along the handle, clear of the
63	body, so the jaws grip the bar — not the rim — and the grip survives
64	the lift. To enable it: (1) author a SECOND perception subgraph that
65	segments the object BODY (e.g. perceiving-object-parts /
66	perceiving-objects with object/subpart name like "the round base of
67	the frying pan"),
68	(2) add `base_obb: OrientedBoundingBox` to THIS subgraph's `inputs`
69	and wire it from that body perception, (3) bind
70	`base_obb = Ref("in.base_obb")` on the `offset_from_base` node.
71	If you do NOT author a body perception, KEEP the `offset_from_base`
72	state but bind only `handle_obb` + `grasp_pose` (omit `base_obb`);
73	the script no-ops and returns the grasp unchanged. NEVER write
74	`Ref("in.base_obb")` unless `base_obb` is a declared subgraph input —
75	that is a graph-validation error.
76	- >
77	When building the collision world, ALWAYS pass
78	`target_mask = Ref("in.target_mask")` AND
79	`target_obb = Ref("in.target_obb")` AND a `target_name`. The mask is
80	the authoritative target silhouette (pixel-accurate); the OBB is a
81	safety fallback. The same `target_name` MUST be passed to `plan`
82	(`plan_grasp.py`) so the planner excludes that mesh from collision.
83	- >
84	ALWAYS include the `approach` state between `offset_from_base` and
85	`observe`. `plan_grasp` disables CuRobo's `use_grasp_approach` flag on
86	the assumption the gripper is already oriented over the target —
87	skipping `approach` makes the descent start from a wrong wrist
88	rotation and slip. `approach_above.py` lifts to a safe height,
89	translates above the target, then rotates in place to the grasp
90	orientation — all three legs route through
91	`robot.go_to_pose_cartesian` (gap.connector.ik.CuRoboBackend), which
92	uses cuRobo's

People who liked this also viewed

6 picks

Type