DynamicDeception

Replication package for the paper:
Dynamic Deception: When Pedestrians Team Up to Fool Autonomous Cars

---

Contents

1. Setup
2. Pedestrians’ Shirts
3. Running the Code
4. Configurations
5. Results

Setup

1. Build CARLA from source
   To customize the environment (stop signs and pedestrians), you must build CARLA from source.
   Follow the official guide:
   https://carla.readthedocs.io/en/latest/build_carla/

   The CARLA version required for replicating our paper is 0.9.15.2:
   https://github.com/carla-simulator/carla/tree/0.9.15.2

2. Install PCLA
   This project depends on the PCLA framework.
   Ensure that:

   • PCLA is cloned on your system
   • The PCLA Python environment is activated
   • You can run PCLA agents

3. Update the PCLA path in main.py
   In main.py, modify line 7 to point to your PCLA directory:

   sys.path.append(os.path.dirname(os.path.abspath("/home/vortex/PCLA")))

4. Replace the Simlingo agent

   A customized agent_simlingo.py script is included in the extras/ folder. This version prints and saves language prompts as .txt files inside the results/ directory.

   Replace the default PCLA Simlingo agent by copying this file to : /path/to/PCLA/pcla_agents/simlingo/agent_simlingo.py

5. Modify the stop sign in Town07

   To replicate our experimental setup, open Town07 in the CARLA UE editor and hide or remove the stop sign at the target intersection used in the paper.

Pedestrians’ Shirts

The pedestrian shirt textures used in our experiments are inside the images/ folder. You may reuse them to replicate the study.
If you want to generate your own patches, please refer to patch_generation folder.

A YouTube tutorial is available showing:

• how to modify pedestrian shirts,
• how to create new pedestrian models,
• and how to spawn customized pedestrians in CARLA.

Running the Code

While the PCLA environment is active, run an evaluation with:

python main.py --patch <option> --scenario <option>

--patch | -p options:

• single – evaluate the single pedestrian attack

• collusion – evaluate the collusion attack

--scenario | -s options:

• static – pedestrian remains still

• dynamic – pedestrian moves with the vehicle

--save-velocities | -sv options:

• Use this flag to flag that you want to save the speed data
• The data will be saved as csv and plot under results/velocity_data/ folder

Configurations

All configuration files are located in the scenarios/ directory.
Each file includes explanatory comments.

general_scenario.yaml

This file contains configuration values shared across all scenarios in a given town.

Important: NPC spawn points change depending on how Town07 is loaded.

• If Town07 is preloaded via the UE editor, use:

  npc_spawn_points: [22, 46, 45, 108]

• If Town07 is loaded through the Python API:

  client.load_world("Town07")

  then you must adjust the NPC spawn points accordingly.

Ego vehicle's spawning point does not need to be changed, as the autonomous vehicle uses coordinate-based spawning (see line ~14 in the YAML file).

Scenario-based YAML files

Each combination of -patch and -scenario has its own YAML file with fine-grained controls such as spawn locations, the movements and etc.

#scenario_evaluation.json

This file defines:

• which PCLA agent performs evaluation
• scenario names
• whether to spawn npc and pedestrians or not
• IDs of the pedestrians being evaluated (must match our scenarios for replication)

Results

All outputs are stored under the results/ directory.

You will find CSV files containing model-level statistics (per-run and aggregated mean ± std). The system-level evaluation were done by analysing recorded videos (used to manually count full stops).

Extras

Additional analysis and benchmarking can be found in the extras/ directory:

• Fisher_Test.ipynb: Contains the implementation and statistical results for our Fisher's test analysis.

• Speed_plot.ipynb: Includes the speed comparison between dynamic and static scenarios in the collusion attack

Project Structure (Overview)

├─ images/ # Shirt textures for pedestrians 
├─ extras/ # Custom agent (agent_simlingo.py) and analysis notebooks
├─ scenarios/ # YAML configuration files 
├─ results/ # CSV logs 
├─ main.py # Python evaluation file 
└─ README.md

Citation

If you find this paper useful, please consider giving it a star 🌟, and cite the published paper