WALKERFACTORY_DECISION.md

Decision record — shipping WalkerFactory with four of six registered walkers

✅ RESOLVED 2026-08-06 — this decision no longer applies

The user re-cooked WalkerFactory in Unreal, keeping only the four shipped walkers (astronaut, boar, deliveryrobot, firefighter) plus CARLA's native 0001–0051, and removing every other imported entry — soldier, wheelchair, caneman, crutcheswoman, cow, tire.

The pack was rebuilt against that clean factory on 2026-08-06. Verified: zero phantom blueprint IDs in the shipped tarball; props and walkers-ccbync byte-identical to the previous build; only walkers-ccby changed (521 bytes smaller). The authors' evaluation build was never modified — the rebuild used a hardlinked staging root — so every published number remains reproducible.

Consequence: the silent-failure path this record describes is gone. A user who installs the pack and runs a route needing a non-shipped prop now gets a loud failure (unregistered blueprint) rather than a quiet one (registered, contentless). The analysis below is retained as the record of why the factory must ship at all.

Status: decided and implemented for v0.9. The open item in §7 is resolved at v1.0. Date: 2026-08-04 · Version stamp: pack v0.9 ↔ arXiv v1

---

1. The problem

The two static props in this pack self-register: CARLA reads Content/<Name>/Config/<Name>.Package.json and the props[] entry creates static.prop.<name>. Dropping the content directory in is sufficient, and nothing in CARLA's base content is touched.

The four walkers do not work that way. Their Package.json also has a walkers[] entry, but that entry is consumed at cook time, not at runtime. At runtime the walker blueprint IDs come from the cooked base-content asset Content/Carla/Blueprints/Walkers/WalkerFactory.{uasset,uexp}. Verified: the string astronaut appears in WalkerFactory.uexp, and /Game/Astronaut/Blueprints/BP_Astronaut appears in WalkerFactory.uasset's import table.

So shipping the four walkers requires shipping WalkerFactory — and the WalkerFactory in the build that produced every published number also registers nine other walkers whose content this pack does not include, including the two non-redistributable benchmark props walker.pedestrian.soldier and walker.pedestrian.wheelchair.

Producing a WalkerFactory that registers only the four would require opening the CARLA UE4 project, editing the factory blueprint and re-cooking base content. Re-cooking in Unreal Engine is forbidden this wave, so that option was not available.

---

2. Options considered

#	Option	Verdict
A	Ship WalkerFactory as-is	CHOSEN. Measured below.
B	Do not ship WalkerFactory; ship walker content only	Rejected — actively dangerous. See §4.
C	Binary-patch the cooked .uasset/.uexp to delete the two entries	Rejected. Requires rewriting a serialised TArray plus the export table offsets, import table and package summary of a cooked UE4 package, with no way to validate short of running it. A corrupt factory registers nothing and fails exactly as silently as the problem it is trying to fix. Not a defensible thing to do by hand.
D	Author stub content for the two absent walkers	Rejected. Authoring a blueprint is UE work.
E	Drop the walkers from v0.9; ship the two static props only	Rejected here, but it is a real option — it contradicts the locked 2026-08-04 decision that v0.9 ships all six, so it is not mine to take. It would cost 72 of the 237 runnable routes. Recorded for completeness.

---

3. Measurement, not inference

I did not want to guess what a factory entry pointing at absent content does, so I measured it. A CARLA build root was assembled out of symlinks to the build that produced every published number (read-only; nothing in it was modified), with Content/ containing only base Content/Carla plus the six shipped directories — i.e. exactly what a user gets after installing this pack over stock CARLA 0.9.15. The CarlaUE4-Linux-Shipping binary was copied (not symlinked) so that UE resolved the project root to the test tree rather than following /proc/self/exe back to the original.

Result, CARLA 0.9.15 server, Town10HD:

WALKER_REGISTERED astronaut     = True     SPAWN OK  extent=(0.213,0.213,0.930)
WALKER_REGISTERED firefighter   = True     SPAWN OK  extent=(0.190,0.190,0.930)
WALKER_REGISTERED deliveryrobot = True     SPAWN OK  extent=(0.432,0.432,0.650)
WALKER_REGISTERED boar          = True     SPAWN OK  extent=(0.470,0.470,0.610)
WALKER_REGISTERED soldier       = True     SPAWN -> None
WALKER_REGISTERED wheelchair    = True     SPAWN -> None

Three things follow, all of them load-bearing:

1. A missing import does not take the package down. The WalkerFactory package loads, and every walker whose content is present registers and spawns normally. Option A does not endanger the four assets we ship.
2. The two absent walkers become "phantoms": present in world.get_blueprint_library(), findable by .find(), carrying the full walker attribute set (gender, age, speed, …) — and unspawnable. world.try_spawn_actor(...) returns None; world.spawn_actor(...) raises RuntimeError: Spawn failed because of invalid actor description.
3. Static props behave differently and better. The four non-shipped props (roadclosedsign, trafficarrowboard, trafficmessageboard, europianarrowboardtrailer) are simply absent from the blueprint library, because props are registered by their own Package.json and there is no shared factory.

---

4. Why option B is worse, and by how much

This is the part that decided it. The relevant code path is CarlaDataProvider.create_blueprint() in scenario_runner/srunner/scenariomanager/carla_data_provider.py:

try:
    blueprints = CarlaDataProvider._blueprint_library.filter(model)
    ...
    blueprint = CarlaDataProvider._rng.choice(blueprints)
except ValueError:
    # The model is not part of the blueprint library. Let's take a default one
    new_model = _actor_blueprint_categories[actor_category]   # 'car' -> vehicle.tesla.model3
    print("WARNING: Actor model {} not available. Using instead {}".format(model, new_model))
    blueprint = CarlaDataProvider._rng.choice(CarlaDataProvider._blueprint_library.filter(bp_filter))

_rng is numpy.random.RandomState, and RandomState.choice([]) raises ValueError (verified), so the except branch is live. Every pedestrian scenario in this benchmark calls request_new_actor(self._pedestrian_blueprint, transform) with no actor_category, and the parameter defaults to "car". Verified at all four call sites (dynamic_object_crossing_modified.py, pedestrian_crossing_stop.py, vehicle_turning_route_pedestrian_modified.py, obstacle_approaching_ego.py).

Therefore:

	What the harness does	Outcome
Option B — walker content shipped, factory not	ID absent from library → filter() → [] → ValueError → substitutes vehicle.tesla.model3 with a stdout warning	A Tesla Model 3 drives at the ego as the "crossing pedestrian". Route completes. Driving Score is plausible. 72 routes × every model silently mis-measured.
Option A — factory shipped	ID present → try_spawn_actor → None → request_new_actor returns None → scenario hits if adversary is None: raise ValueError("Failed to spawn pedestrian")	The route errors out and is recorded as failed.

Option B converts our four good assets into silent Tesla substitutions. Option A leaves two IDs that fail noisily. That is not a close call.

---

5. What we are actually shipping, and its precise consequence

ood-perceptionbench-walkers-ccby-v0.9.tar.gz contains CarlaUE4/Content/Carla/Blueprints/Walkers/WalkerFactory.{uasset,uexp}, taken byte-for-byte from the build that produced every published number (sha256(WalkerFactory.uexp) = f5f418f1b39a0100dd1d33bebf3fcc0566ae2ee2f607995c2c5c3fb94045957c).

Consequence, stated precisely:

Installing this pack overwrites one file in CARLA's base content. Afterwards the blueprint library advertises nine walker blueprint IDs whose content is not in the pack: soldier, wheelchair, ball, caneman, cow, crutcheswoman, deer, labrador, tire. All nine are unspawnable: try_spawn_actor returns None and spawn_actor raises. They cannot silently produce a wrong prop, but they can waste your time if you try to use them. Two of them — soldier and wheelchair — are referenced by 36 of the benchmark's 475 route XMLs, and those 36 routes are not runnable at v0.9. The other seven belong to unrelated experiments and appear in no benchmark route.

Because this file overwrites base content, the pack is hard-locked to CARLA 0.9.15. Do not install it over any other CARLA version.

tools/verify_pack.py asserts all of this — that the six shipped IDs spawn, and that every phantom does not.

---

6. The clean fix, and why it is nearly free

The right artefact is a WalkerFactory registering only the walkers actually shipped. That needs one UE editor session: open the CARLA project, delete the unwanted rows from the factory blueprint's parameter array, re-cook base content.

A batched Unreal Engine session is already on the v1.0 critical path (replacing the two unlicensable walkers, refused-vehicle swaps, the vehicle.ood.* rename and its VehicleFactory re-cook, plus a WalkerFactory re-cook for the replaced walkers). Trimming the factory is a few minutes inside a session that must happen anyway. Recommend folding it in rather than scheduling separate work. After that session the phantom list should drop to zero and this document can be retired.

---

7. Open item

This is a known, accepted defect of v0.9, not an oversight: the pack ships a WalkerFactory that advertises nine unusable blueprint IDs, because the alternative (§4) silently substitutes a Tesla Model 3 for a pedestrian across 72 routes. It is scheduled to be fixed in the v1.0 UE session described in §6, after which the phantom list should be empty and this document can be retired.

---

Appendix — incidental findings from this investigation

A. The vehicle.ood.* rename has not reached the cooked VehicleFactory. Content/Carla/Blueprints/Vehicles/VehicleFactory.uexp still contains Tohoku, Inkas, Caterpillar, Hamm, Sedane, Hatchback, DumpTruck, RoadRoller. The ood.* rename landed in the route XMLs, the route generator and vehicle_classification.json, but the blueprint IDs a running CARLA actually serves are still the old trademarked ones. This is consistent with the fact that a vehicle rename needs a full re-cook, scheduled for the v1.0 UE session) and is not a v0.9 blocker — no vehicle asset ships at v0.9 — but it does mean the shipped route XMLs reference six vehicle blueprint IDs that exist in no current build. The frozen route set and the acceptance goldens both expect that.

B. Content/Boar/Animations/ contains 60.3 MB of uncooked editor assets. Seven .uasset files with no .uexp sibling (alerted, digging_feeding, observing, sniffing, trot, wake_up, wound), dated 2025-12-09 rather than the 2026-07-07 cook. The shipping runtime cannot load them and no cooked package imports them — the boar's locomotion uses walk/run/sleeping via BS_Boar and laydown via BP_Boar, all four of which are properly cooked and shipped. Excluded from the pack; hashes recorded in build/EXCLUSIONS.tsv.

C. Content/Firefighter/Blueprints/ has a case-insensitive filename collision. BP_Firefighter.uasset and BP_FireFighter.uasset differ only in one letter's case. On Linux both exist and BP_Firefighter_v2 correctly inherits from the lower-case one; on macOS or Windows one silently overwrites the other during extraction, and the surviving file's internal package name may not match the path UE looks it up by — which would take the firefighter down. The upper-case BP_FireFighter is referenced by nothing (that exact string occurs in exactly one file in the whole build: itself), so it is excluded from the pack and the collision is gone.