Recording and Results

Use the same scenario, model, seed/configuration and run length when comparing experiments. A successful process exit alone does not establish perception accuracy or attack success.

Bounded Runs and Logs

In the OpenCDA container shell, with CARLA_HOST set as in Connect from OpenCDA:

python opencda.py -t v2xp_datadump_town06_carla \
  --carla-host "$CARLA_HOST" --ticks 200

The default evaluation directory is simulation_output/evaluation_outputs/<scenario>_<timestamp>/. It contains opencda.log.json and the reports/plots produced by enabled metrics. Scenario names containing / create nested output paths. Use --log-file /path/to/run.log.json to override the log location.

--verbose 1, 2, and 3 select warning, informational and debug logging respectively.

Profiling Slow Runs

--profile measures how much time Python spends in the scenario’s functions. Use it to investigate a slow simulation. It adds profiling overhead, so leave it off when measuring normal simulation performance.

For example, in the OpenCDA container shell:

python opencda.py -t v2xp_datadump_town06_carla \
  --carla-host "$CARLA_HOST" --ticks 200 --profile

The runner saves profile_output.prof in that run’s evaluation directory. This is a binary Python cProfile report, not a video or a perception metric. To inspect it, replace RUN_DIRECTORY with the directory printed for your run:

python -m pstats RUN_DIRECTORY/profile_output.prof

At the pstats prompt, enter sort cumulative followed by stats 20 to see the 20 functions with the highest cumulative time (including calls to other functions). Enter quit to exit.

Recording

--record enables CARLA recording and scenario sensor data dumping:

python opencda.py -t v2xp_datadump_town06_carla \
  --carla-host "$CARLA_HOST" --record --ticks 200

CARLA recorder files are written by the CARLA server, so their location belongs to its container on Linux or its host process on Windows. Copy files out of a container before removing it if they are not on a bind mount.

Cooperative Perception Output

Add these flags to a cooperative-perception or AdvCP command:

Flag	Result
--save-vis	Frames under simulation_output/coperception/vis_<mode>/<scenario>_<timestamp>/. Available views depend on visualization configuration.
--save-npy	Prediction and ground-truth arrays under simulation_output/coperception/npy/<scenario>_<timestamp>/.
--show-video-vis	Live visualization. Requires a working display in the container.

These directories are inside the mounted OpenCDA checkout and remain on the host when the container is recreated. Utility Scripts describes make_video.py for converting a saved frame directory to a video.

Metrics and Warmup

The scenario’s coperception.metrics configuration selects metric behavior. Use its warmup settings when interpreting short runs. Inspect the logged agent IDs and compare per-run metrics alongside the rendered output. Missing attackers, empty inputs, or a target outside the sensor range can make an experiment uninformative even when the simulator continues running.