perceiving-deformable-linear-objects

Read the cable's centreline once per step, and carry it through the loop as text.

Why this is its own skill

A rigid object is perceived once and then transformed; a cable has no pose, only a shape, and the shape changes every time a crossing moves it. The thing a routing loop needs is therefore an ordered 3D centreline — 42 nodes in arc-length order, not a bounding box and not a principal axis — and a rule for when to trust a fresh reading of it over the one it already has. That rule is the skill: a cold fit has no history and is the better estimate whenever the whole rod is in frame; a tracker is the better estimate only when it is not. Which case a frame is in is measurable (length against the prior) and the script routes on it.

Measured over eight arm poses from a bench camera pitched 55 degrees off the vertical, scored against the rod's own bodies:

estimate	median	p90	max
cold fit per frame	6.7 mm	8.4 mm	9.3 mm
tracked from the prior	9.0 mm	9.7 mm	9.9 mm

The tracker is the worse estimate when nothing is hidden, and monotonically so across passes (7.6, 8.6, 9.2, 9.6, 9.6, 9.9 mm): each pass drags the prior along instead of carrying a hidden stretch through. It is still the right fallback, because the cold fit's own failure is sharp rather than gradual — a skeleton broken by the arm merges to one fragment, measured at 338 mm of a 488 mm rod (0.69), while honest fits ran 0.99–1.06 of the truth. The whole_rod gate (0.80) sits between those two populations.

When to use

• A loop needs the current shape of a cable, rope, or hose once per station or per step, and the shape only changes when the robot moves it.
• A fixed third-person RGB-D camera can be placed where it sees the whole object. Measured: an eye-in-hand camera frames a fifth of the cable at best, and an A/B over three episodes came out identical (0.667 mean) — the move it costs buys nothing. Pitch the camera off the vertical: overhead, the hand covered the rod exactly when the rod was moving (28 of 42 nodes visible, 60 mm RMS); at 55 degrees it sees 41 of 42 and fits at 6.7 mm.

When NOT to use

• Every-frame tracking during a motion. This skill spends one detector call and one fit per read; call it between motions, not inside one.
• Rigid, linear parts (a shaft, a handle) — fit an axis with geometry.fit_linear_feature instead.

State flow

prior == "" or index == 0 ──► seed from scene.rod ──► found (seeded)
           │ no seed
           ▼
prior == "" ──► cold fit every mask candidate, take the THINNEST ──► found (initialised) | lost
           │ prior present
           ▼
for camera in (chosen, track_camera, init_camera):
    cold fit ──► length ≥ whole_rod × len(prior)? ──► found (refit)
    else track prior onto the frame ──► ≥ min_visible nodes seen? ──► found (tracked)
nothing improved ──► found (held; rod == prior)

1. Seeded. On the first pass (index == 0) — or whenever there is no prior — the survey's rod (the scene JSON's rod key, from perceiving-routing-fixtures) is returned as the model. The model must be born from an unobstructed view of the whole object, and that view exists exactly once per episode, before the first move; re-fitting after the hand has parked over the first station read 381 mm of a 500 mm rod.
2. Initialised. With no prior and no seed, the bench camera is read, up to mask_candidates masks are fitted, and the thinnest plausible one wins. The top-scoring mask is not always the cable: on a grey bench with white arms a prompt sometimes returns the rod merged with an arm, scored confidently — measured on a 700 mm rod as 963 and 983 mm seeds on two of six episodes, both of which failed. Radius (mask area over centreline length) separates them where score cannot; max_rod_radius (12 mm) is the ceiling.
3. Refit. The look's camera is tried first, then track_camera, then init_camera. A fresh fit whose arclength reaches whole_rod of the prior's is taken as is.
4. Tracked. A short fit hands the frame to curve.track_centerline, which moves the prior's nodes by what the frame says about each and carries the unanswered ones on the displacement field of their visible neighbours. The update is believed only when at least min_visible nodes (6, about 70 mm of a 500 mm rod) had real correspondence — a tracker that accepted every frame would walk the model onto whatever happened to be visible.
5. Held. Nothing improved on the prior; it is handed back unchanged.

Inputs

• prior — the centreline the loop carries, as JSON text; empty on a cold start. The graph binds it to the latest rod written upstream (the survey's on the first pass, this skill's own afterwards).
• scene — the survey JSON from perceiving-routing-fixtures; only its rod is read here.
• camera — the camera the look chose for this pass (empty: the bench camera).
• index — the pass number; 0 is the pass before the first move.
• Bench constants as parameters with the measured defaults: init_camera and track_camera ("cable"), rod_query ("thin white cable"), rod_score (0.20), whole_rod (0.80), min_visible (6), max_rod_radius (0.012), curve_nodes (42), mask_candidates (4).

Outputs

• rod — the centreline as JSON text: the points of a Centerline, an ordered list of [x, y, z] metres at 0.01 mm precision. It is passed as text so it round-trips through a loop unchanged (json.loads gives the list; a JSON Centerline object with a points key is also accepted on the way in). On held and lost it is the prior that was given ("null" when there was none), so a pass never binds a worse model than it had.
• arclength_m, source (seeded | initialised | refit | tracked | held | no-camera | no-mask | no-fit), camera (which view carried this update), visible (nodes with real correspondence; the node count for a fit).

The router field is route. A node that raises binds no outputs, so the next reader of rod falls back to whatever was written before it.

Required end states

End state	Meaning
found	rod carries a centreline (see source for how it was earned).
lost	No model could be born and there was no prior to hold.