This is the replication package of the ACM TOSEM paper "Finding Safety Violations of AI-Enabled Control Systems through the Lens of Synthesized Proxy Programs", co-authored by Jieke Shi, Zhou Yang, Junda He, Bowen Xu, Dongsun Kim, DongGyun Han, and David Lo. It contains the source code for applying our tool (Synthify) and the baseline (PSY-TaLiRo) on the falsification of AI-enabled control systems. The plant of each control system is in envs and the AI controllers are in checkpoints. Moreover, the hyperparameters of each control systems are in config.json.
To reproduce our experiments, we provide a Dockerfile to help build the experimental environment. Please run the following scripts to compile a docker image:
docker build -t YOUR_CUSTOM_TAG .
Then, please run the docker as a container, and mount your code folder to a folder in your container.
docker run -it -v YOUR_LOCAL_REPO_PATH:/root/Synthify YOUR_CUSTOM_TAG
Given the same number of falsification trials, i.e., 50 trials, we provide two bash scripts to run Synthify and PSY-TaLiRo on the four AI-enabled control systems, respectively. Note that the script will repeat experiments for 10 times on each control system. Please run the following scripts to reproduce the results of Synthify:
bash synthify.sh
It contains the following command for each control system:
python3 synthify.py --env=ENV
Please run the following scripts to reproduce the results of PSY-TaLiRo:
bash psy_taliro.sh
It contains the following command for each control system:
python3 psy_taliro.py --env=ENV
We strongly recommend users to run the above bash scripts, as they help save the results of each trial in a separate file for subsequent analysis. Otherwise, results of each trial will be printed on the screen, and users need to manually save them for subsequent analysis.
Parse results: After saving the results of each trial into results folder, please run the following command to parse the results:
python3 parse_results.py
The parsed results will be printed on the screen with average results and standard deviation, as well as significance test results with effect sizes.
Since we report the average results of 10 experiments in the paper, and the falsification algorithms are stochastic, users may get slightly different results. Please note that such results usually can be tolerated, i.e., they mostly do not conflict with the conclusions of the paper. We also provide the code for significance test results with effect sizes to help users analyze the results.
If users want to train the AI controllers by themselves, please run the following command:
python3 DDPG.py --env=ENV
But we do not recommend users to do so, as it may take a long time to train the AI controllers. We have provided the trained AI controllers in checkpoints folder, and we truly thank the authors of An inductive synthesis framework for verifiable reinforcement learning for providing the trained AI controllers in the repository of VRL.
If you have any questions, please feel free to contact us at [email protected].
If you use our code, please cite our paper:
@article{10.1145/3715105,
author = {Shi, Jieke and Yang, Zhou and He, Junda and Xu, Bowen and Kim, Dongsun and Han, DongGyun and Lo, David},
title = {Finding Safety Violations of AI-Enabled Control Systems through the Lens of Synthesized Proxy Programs},
year = {2025},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
issn = {1049-331X},
url = {https://doi.org/10.1145/3715105},
doi = {10.1145/3715105},
note = {Just Accepted},
journal = {ACM Trans. Softw. Eng. Methodol.},
month = jan,
keywords = {Falsification, Search-based Testing, AI-enabled Control Systems, Program Synthesis}
}
The paper is available at https://dl.acm.org/doi/abs/10.1145/3715105.