anchoring-a-free-end

SkillDev tools

Pins one end of a cable with the support hand before any crossing

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the anchoring-a-free-end skill

What this skill tells your AI

The instructions your AI receives, as published by graph-robots/open-robot-skills in skills/anchoring-a-free-end/SKILL.md and read by ahel’s review.

Take the cable's far end with the support hand, before any crossing.

Why this is its own skill

Why it matters is a property of the material, not of the plan. A cable held at one end and pulled sideways bends; a free cable translates. Without an anchor the first crossing hauls the whole cable toward one station and the next has nothing left to grasp — measured, 26 of 42 segments piled at a single station. Every later crossing then works on a cable that has already moved out from under the model that was measured.

Failure is fatal here in a way a missed look is not. A weave without an anchor is not a degraded weave, it is a different and much worse task, so the skill routes unanchored for a graph to abort on rather than carrying on and reporting the result as a policy failure.

What it needs from the connector

This skill requires a connector that registers cable.plan_support, a privileged support-hand planner that reads the true rod state and returns the anchor grasp as joint-space legs (stages), the hand that owns them (arm_id, arm_name), the close width (close), the end joint state (q_end), the grasp orientation (quat), the rod segment taken and the IK error of the worst waypoint (worst_grasp_error) — in simulation the cable-routing suite's connector provides it. It also needs robot.set_grip (a ramped jaw command in metres, addressed by arm_id) and robot.wait_steps (hold still while the scene runs on). The generic connector registers none of the three, which is why the bundle is sim-only today.

Recommended subgraph state flow

anchor_far_end ──anchored──▶ (the first crossing)
       └──unanchored──▶ abort
  1. Ask cable.plan_support(approach=) for the grasp.
  2. Refuse when there are no legs or worst_grasp_error exceeds 50 mm — past that the chain is in a different basin and what streams is a whip with the cable near the jaws.
  3. Stream each leg with robot.execute_trajectory on the plan's arm_id.
  4. robot.set_grip(width_m=plan.close, arm_id=plan.arm_id), then robot.wait_steps for the jaws' own travel: a finger command is a target, not a state, and the first crossing must not start while the pads are still moving.
  5. Return the solved chain (q_support) and grasp orientation (quat) as JSON lists: every later pay-out seeds its IK from them, and re-deriving them per crossing would let the support chain drift into a different basin mid-weave.

Inputs and outputs

The script takes one parameter, approach (m above the grasp the enter leg starts from; 0.055). Outputs: segment (the rod segment held; -1 when refused), worst_error_m, and q_support / quat as JSON strings ("null" when refused).

What a CPU test can and cannot exercise

The refusal gate and the call order (plan, stream each leg, close on the plan's width and hand, dwell) run against a canned plan. What a CPU test cannot exercise is the planner behind cable.plan_support — whether the legs it returns reach the rod — nor whether the pinch actually holds under the first crossing's pull.

Required end states

End stateMeaning
anchoredThe far end is held; segment, q_support, quat bound.
unanchoredRefused before anything moved. Abort the weave.

Signals

GitHub stars
41
Forks
7
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
anchoring-a-free-end
Source
github.com/graph-robots/open-robot-skills