!! Repo under construction !!

This directory contains the basic setup for running Scenic scenarios in CARLA to create and store images for scenes based on scene graphs.

• run_scenic.py: This script executes the scenic scenarios and captures data.
• *.scenic: These files define the specific scenarios (e.g., obstacle_clear.scenic, obstacle_rain.scenic, etc.).
• certificate.py: This script runs certification on the generated images using VQA models (Qwen or LLaVA).

1. CARLA Simulator: You need a running CARLA server.
   • Download CARLA (e.g., 0.9.14 or newer).
   • Run the server: ./CarlaUE4.sh (defaults to localhost:2000).
2. Scenic: Install the Scenic language.
   • pip install scenic
3. Python Packages:
   • carla: Ensure the Python client matches your server version.
   • transformers, torch, Pillow, statsmodels (for certification).

1. Start CARLA server.
2. Run the experiment script:

   python run_scenic.py

To certify the results using a VQA model:

python certificate.py --question "Is there a car?" --answer "yes" --image_dir "path/to/images" --model_name "qwen"

1. Brake with an obstacle very close

   • Scenario Files: obstacle_*.scenic
   • Question: Answer with only ‘Yes’ or ‘No’: Is it necessary for the vehicle to stop here?
   • Correct Answer: Yes

2. Building-route blockage check

   • Scenario Files: obstacle_*.scenic
   • Question: Answer with only ‘Yes’ or ‘No’: Does a building block the vehicle’s route?
   • Correct Answer: No

3. Right-turn feasibility

   • Scenario Files: right_*.scenic
   • Question: Answer with only ‘Yes’ or ‘No’: Is a right turn physically possible here?
   • Correct Answer: Yes