graph-robots seating-a-cable-crossing
Seat one crossing, then look, and hand on an honest number.
Why this is its own skill
A weave that streams its legs and hopes never finds out which crossing failed.
The whole difference between a full solve and a third of one on the weave
bench was one crossing that did not take and was never looked at again. This
skill closes that loop for a single crossing: stream the planner's place
leg, wait for the rod to come to rest, then segment the rod, fit its
centreline and count the densified curve inside the station's seat box.
"Did not take" is a fraction, not a touch. The task scores containment as
the share of the rod inside a box a few centimetres wide beside the station,
on the side the route requires (>= 0.047 on the weave bench). Measured on the
failures, a losing station is almost never on the wrong side: it is on the
right side with too little of it there.
What it deliberately does not do
It measures and does not correct. The obvious correction — push the hand further onto the required side when the box is short — was built and measured and made the seat worse (0.016 to 0.000 after a 14 mm nudge): the cable is not short of the box, it is crossing it at an angle, and driving the contact point further in rotates it further. A correction wants a different move, arriving along the seat rather than across it, and that belongs in the carry planner.
Routing
The script always returns route: done; the verdict is in the outputs.
seated is fraction >= seat_fraction; fraction is 0.0 when the seat
could not be measured at all (no camera frame, no depth, no rod mask, a fit too
short to be a curve) and detail says so. A graph that wants to branch on the
verdict reads seated, not the exit. blocked is the exit a subgraph declares
as on_error for a tool that raises during the place or the look.
Which station
The plan's station field names the station the planner actually worked, and
that — not the loop's pass index — selects the seat box. The planner does not
work stations in pass order (far-side crossings first, near-side seats last,
each from the anchored end outwards), so indexing the perceived list by the
pass number measured a different station's box from the one just worked on
every pass. index is the fallback only when the plan carries no station.
Inputs and parameters
| Input | Shape |
|---|---|
plan |
JSON: {"stages": [{"stage": "place", "trajectory": Trajectory, ...}], "station": int, "side": ±1} from the crossing planner |
scene |
JSON: {"stations": [[x, y], ...], "sides": [±1, ...]} from the fixture survey |
index |
the pass number; used only when the plan names no station |
arm_id |
the work arm that streams the place leg |
Parameters with the weave bench's numbers as defaults: seat_fraction (0.047),
seat_inner / seat_outer (0.003 / 0.045 m out from the station on the
required side), seat_half_y (0.035 m along the station line), camera
("cable"), rod_query ("thin white cable"), curve_nodes (42). Read a
different bench's off its own task document rather than reusing them.
grip_open_m / grip_close_m are accepted for uniform binding and unused.
What a CPU test can and cannot exercise
The fraction arithmetic (densification, the side sign, the box bounds) and the
plan-station indexing run against canned observations and a canned centreline
fit. What a CPU test cannot exercise is the perception behind
sam3.segment_text and curve.fit_centerline on a real frame, and whether a
streamed place leg actually lays cable where the plan said.
Required end states
| End state | Meaning |
|---|---|
done |
Streamed and measured; read seated / fraction / detail. |
blocked |
A tool raised; the subgraph's on_error. |