TASK 31 — Ground truth for all fifteen scenes, and what it refuted

Half the official question set could not be scored offline, because viz/data/<scene>.json existed for seven scenes out of fifteen. TASK 30 listed fixing that as one of two next steps and priced it as a large job.

It was not a large job. The estimate was wrong in a way worth recording, because the same mistake would have deferred it again.

The export had two halves and only one was expensive

export_viz.py writes one manifest from two sources:

half	source	cost
perception	a recorded tour under frames/<scene>_tour	hours of sim time per scene
ground truth	object_list.txt + map.ply inside the scene model	seconds, reads a zip

They were welded together by a single early return: no tour, no export. So the cheap half inherited the expensive half's price, and eight scenes had neither. All fifteen scene models were already on the laptop the whole time.

Now a scene with no tour exports ground truth only. The eight ran in about a minute, on the laptop, with the sim host idle:

scene	gt objects	scene cloud
home_building_1	432	9.5 M → 500 k
home_building_2	227	5.1 M → 500 k
hotel_room_1	86	7.6 M → 260 k
hotel_room_2	95	4.9 M → 146 k
livingroom_1	106	7.4 M → 262 k
livingroom_2	88	7.4 M → 263 k
livingroom_4	120	5.3 M → 168 k
studio	73	6.7 M → 232 k

Three other things had to change with it:

• Several roots. The models arrived as three download batches, two of whose directories have a space in the name, and the sim host keeps them unpacked instead. --gt-root now takes a list and reads a zip or an unpacked directory, rather than requiring a fourth copy that stays in sync with both.
• A guard against self-harm. The corpus lives on whichever machine recorded it, so running --scene all on the laptop finds no tours at all — and would have rewritten the seven good exports as ground-truth-only, destroying work that cannot be reproduced here. An existing export with frames is now kept unless --force.
• The index is the union. Exporting one scene used to rewrite index.json to that scene alone, hiding the rest of the page's data.

The viewer tolerates a scene with no tour: no scan, no detections, no robot, the camera framed on map.ply instead of on a lidar cloud it does not have. Checked in the browser on studio — 65 boxes and the scene cloud draw, and office_1 still replays frame-by-frame.

What it confirmed

Anchor size. TASK 30 concluded that anchor size is loft's problem rather than a general one, from 44 anchors over 7 scenes. The sample is now 92 over 15, and the distribution barely moves:

subtends at start	7 scenes	15 scenes
0–2°	5 %	3 %
2–5 °	11 %	10 %
> 5°	84 %	87 %

The conclusion survives doubling the sample. The new hard case is home_building_1's clock, 0.20 m at 8.8 m — 1.3°, smaller than loft's remote.

How near is near. traj_tolerance.py now measures 136 named-object distances instead of 63: median 0.59 m against the 0.58 m TASK 30 quotes, 81 % within a metre, and unnamed objects still sit at 1.95 m. The target we drive at did not move.

The questions are answerable against this ground truth. 74 of the 75 questions name at least one object using the scene's own label. The single miss is chinese_room's "count the number of chairs with pillows on them" — plural morphology against a chair label, not a missing object.

What it refuted

The third keep-out question was the one without ground truth. It is livingroom_2 q5, and it breaks two things.

The questions are not written against the scenes' vocabulary. The sentence says "avoiding the path between the TV and the tea table". livingroom_2 has no tea table; it labels that furniture coffee table. chinese_room's question uses the same two words against a real tea table label. So an anchor name cannot be looked up literally, in either direction.

KEEPOUT_M = 1.2 is refuted. With the third case measurable, the bracket keepout_radius.py reports no longer brackets anything:

lower bound	0.86 m	our p90 centre error + robot half-width + slop
upper bound	0.81 m	tightest official clearance (livingroom_2 q5)

There is no single half-width that is both safe for us and legal for the reference path.

Worse, the loop does not build a corridor at all — bind_constraints puts a 1.2 m disc on each anchor object. Measured against the official paths:

anchor	official path passes	1.2 m disc
chinese_room chair[2]	0.59 m	forbids it
chinese_room chair[3]	0.71 m	forbids it
livingroom_2 coffee table[0]	0.81 m	forbids it
chinese_room chair[0]	1.27 m	ok
chinese_room chair[5]	1.98 m	ok

3 of 11 anchor instances. In chinese_room whether it fires depends on which of six chairs the model binds; livingroom_2 has exactly one coffee table and would fail outright.

Measuring to the box surface rather than the centre does not rescue it — the official paths pass 0.22 m, 0.29 m and 0.30 m from those boxes, roughly the robot's own half-width.

The model is wrong, not the number. "Avoid the path between A and B" forbids crossing the gap, and the reference paths brush right past A while respecting the corridor. A disc on A forbids the wrong thing. keepout_radius.py already measured the corridor geometry; ConverterModel(keepout=[(xy, r)]) implements discs. The measurement script and the implementation have disagreed about what the constraint is since the constraint was built.

No radius was changed. Picking one is a design decision — corridor versus disc, and what to do when no radius satisfies both bounds — and the keep-out path has never worked end to end anyway (TASK 30: one live anchor, lifted onto the ceiling). It is recorded here rather than patched.

State

viz/data/ holds all 15 scenes (gitignored; 160 MB). Changed: scripts/export_viz.py, scripts/keepout_radius.py, viz/app.js. Uncommitted, along with everything from TASK 30.

Open

• gate — still the largest single item: 10 of 30 instruction questions, parsed but not enforced. Now measurable on all 15 scenes.
• avoid as a corridor, per above; 3 questions.
• Recorded tours for the 8 new scenes, if we want perception numbers there rather than ground truth alone. This is the expensive half.
• Repeat runs per scene, unchanged from TASK 30.