fix: four paths that skipped a rule FIXS already states - #324
Open
yunlishao-vibe wants to merge 4 commits into
Open
fix: four paths that skipped a rule FIXS already states#324yunlishao-vibe wants to merge 4 commits into
yunlishao-vibe wants to merge 4 commits into
Conversation
… argv -DisableFrameTraceCapture went into EDITOR_LAUNCH_FLAGS and reached the three UE4Editor launches FIXS assembles itself - run_cosim, place_tls, place_signs. The map cook is a fourth, and FIXS does not build its command line: it shells out to CARLA's Util/BuildTools/Import.py, which assembles the ImportAssets commandlet argv from its own list. So the flag never arrived there and the "Locate main RenderDoc executable..." dialog still blocked a cook, the one launch with no timeout at all. UE4 appends whatever UE-CmdLineArgs holds to any command line it parses (Engine/Source/Runtime/Core, LogSuppressionInterface.cpp:634), so setting it in the child environment puts the flag into an argv owned by someone else without patching their script. Appends rather than assigns, so an operator who already uses the variable keeps what they set.
carla.json names one interpreter and every entry point re-execs under it, but naming the interpreter does not settle which packages it imports. A conda env has no pyvenv.cfg, so CPython leaves ENABLE_USER_SITE on and puts the per-user site directory AHEAD of the env's own site-packages. One `pip install --user` therefore shadows an env-installed package in every env on the machine at once. Measured: a carla built from a 0.9.16 source tree landed in %APPDATA%\Python\Python310\site-packages and won over the 0.9.15 wheel the config had installed. The client connected to a 0.9.15.2 server, CARLA only WARNED about the version gap and connected anyway, and the run died on the first camera image inside libcarla - ImageTmpl.h's GetWidth() * GetHeight() == size() assertion, a C++ assert, so it took the process down instead of raising something python could report. Two moves, because an env var cannot repair a process that has already booted: export PYTHONNOUSERSITE so every child starts without the directory, and drop it from this process's sys.path for the case where we are already on the configured interpreter and so never re-exec. Done at import rather than from a call each entry point must remember, because it has to beat `import carla` on every path - including run_cosim --version, which returns above the re-exec and would otherwise fingerprint the shadowed copy as "carla present", in the output whose own help text calls it what to paste into a bug report. The notice names the interpreter, not the directory that was dropped: which python you got is the question a shadowed import makes unanswerable, and it is the only part of this a reader can act on. It prints only on the paths that return - a re-exec already names the python it switches to.
Every tl_table FIXS writes is in SUMO coordinates. Measured across five cooked
maps on one machine - uga, roosevelt, mlk (two vintages), atlanta: 1156 rows,
not one negative y. CARLA's world is left-handed, so a SUMO point (x, y) is
CARLA (x, -y), and _table_junctions already knows this - it just did the flip
only when --no-net-offset was passed.
Off was never right for a table FIXS generated. It is right only for a table
already written in CARLA coordinates, which is what the new --net-offset is
for. Without the flip the spectator is sent to +y where there is no road: on
the UGA campus map the busiest junction sits at y=1200, so the camera framed a
point 2400 m north of it and the map read as having failed to load - the same
symptom _frame_from_map was added to cure.
The flag is now tri-state so the per-map setting can still opt out, and an
explicit flag finally beats it. The old expression was
args.no_net_offset or settings.get("net_offset") == "zero"
which meant that on a map recording "zero" the flip could not be turned off
from the command line at all, despite the comment two lines above promising
that explicit CLI flags override. "zero" keeps meaning ON, so existing per-map
settings carry over unchanged; a map opts out with net_offset: "keep".
The setup is saved before the LAUNCH, deliberately - a run that failed to start is the one you want to come back to and tweak. But it is saved after the map IMPORT, and the import is the step that fails for reasons outside this script: a cook that crashes the editor, a bundle that will not open. A failed cook therefore discarded the app, map and local pick just chosen, and the next run asked the whole questionnaire again - at exactly the moment you want to change one answer and retry. save_partial writes the record as soon as the questionnaire's answers are settled, tagged with the stage that had not run yet, under the same name - so finishing later overwrites it instead of leaving a second entry behind, and the full save clears the tag. config/config_scope are not checkpointed: they are resolved after the import, and the record already carries what the previous run knew, which is the right thing to come back to. The stage is a name, not a flag, because it is shown in the picker: "map import" reads as a thing to retry where a bare "incomplete" reads as corruption. A headless run says it out loud, since it draws no list and silently replaying a setup whose map never cooked is how the failure gets rediscovered further downstream.
yunlishao-vibe
force-pushed
the
fix/cosim_setup_gaps
branch
from
August 26, 2026 15:29
2e65515 to
5dce66b
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four independent fixes, one commit each, all found while bringing a new
RoadRunner map (UGA campus) up on a source build. They share a shape: FIXS
states a rule, and one code path did not enforce it.
fix(#311)— the cook never got the editor flags-DisableFrameTraceCapturereaches the three UE4Editor launches FIXS assemblesitself. The map cook is a fourth, and its argv belongs to CARLA's
Util/BuildTools/Import.py. Observed command line from a failed cook:No flag. Fixed by setting
UE-CmdLineArgsin the child environment, which UE4appends to any command line it parses.
fix— per-user site-packages shadowed the configured envcarla.jsonnames one interpreter; that does not settle which packages itimports. A conda env has no
pyvenv.cfg, so CPython puts the per-user sitedirectory ahead of the env's own
site-packages, and onepip install --usershadows an env-installed package in every env at once.A 0.9.16
carlabuilt from a source tree won over the 0.9.15 wheel the configinstalled. CARLA only warns on a version gap, so the run connected to a
0.9.15.2 server and died on the first camera image inside libcarla —
ImageTmpl.h'sGetWidth() * GetHeight() == size()assert, which takes theprocess down rather than raising into python.
Exports
PYTHONNOUSERSITEfor children and drops the directory from thisprocess's
sys.path. Done at import, so it also coversrun_cosim --version,which returns above the re-exec and was reporting the shadowed copy as
carla present.fix— the TL table is in SUMO coordinates, so flip it by defaultMeasured across five cooked maps (uga, roosevelt, mlk ×2, atlanta): 1156
rows, not one negative y. CARLA's world is left-handed, so SUMO
(x, y)isCARLA
(x, -y). Without the flip the spectator went to+ywhere there is noroad — on UGA the busiest junction is at
y=1200, so the camera framed a point2400 m north of it and the map read as failed-to-load.
The flag is now tri-state, and an explicit CLI flag finally beats the per-map
setting. The old
args.no_net_offset or settings.get("net_offset") == "zero"meant the flip could never be turned off on a map recording
"zero", despitethe comment above it promising the opposite.
"zero"still means ON; a mapopts out with
net_offset: "keep".feat— checkpoint the setup before the map importThe setup is saved before the launch, but after the import — and the import
is what fails for reasons outside the script. A failed cook discarded the app,
map and local pick just chosen, and re-asked the whole questionnaire at exactly
the moment you want to change one answer and retry.
save_partialwrites the answers early under the same name, tagged with thestage that had not run; the full save clears the tag.
Verification
py_compile, all four files) —bisectable.
carlaresolves from the user directory before and from the env's0.9.15 wheel after, and a child process inherits the change.
--no-net-offset,--net-offset} × {no setting,"zero","keep"}.a child on Windows.
entry, with a normal record never gaining the key.
Not yet exercised end to end in a live co-sim; each change was verified in
isolation.