Anchoring A Free End

anchoring-a-free-end

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